diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/__init__.py b/build/torch26-cxx11-cu126-aarch64-linux/moe/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..0d9d8227711def6d51c24b5fd779a3d13681d88c --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/__init__.py @@ -0,0 +1,80 @@ +import torch + +from . import layers +from ._ops import ops +from .fp8_utils import per_token_group_quant_fp8, w8a8_block_fp8_matmul +from .fused_marlin_moe import fused_marlin_moe +from .fused_moe import fused_experts, fused_moe, fused_topk, grouped_topk +from .scalar_type import ScalarType, scalar_types + + +def gptq_marlin_moe_repack( + b_q_weight: torch.Tensor, + perm: torch.Tensor, + size_k: int, + size_n: int, + num_bits: int, +) -> torch.Tensor: + num_experts = b_q_weight.shape[0] + assert size_k % 16 == 0 + output = torch.empty( + (num_experts, size_k // 16, size_n * (num_bits // 2)), + device=b_q_weight.device, + dtype=b_q_weight.dtype, + ) + for e in range(num_experts): + output[e] = ops.gptq_marlin_repack( + b_q_weight[e], perm[e], size_k, size_n, num_bits + ) + return output + + +def awq_marlin_moe_repack( + b_q_weight: torch.Tensor, + perm: torch.Tensor, + size_k: int, + size_n: int, + num_bits: int, +) -> torch.Tensor: + num_experts = b_q_weight.shape[0] + assert size_k % 16 == 0 + output = torch.empty( + (num_experts, size_k // 16, size_n * (num_bits // 2)), + device=b_q_weight.device, + dtype=b_q_weight.dtype, + ) + for e in range(num_experts): + output[e] = ops.awq_marlin_repack(b_q_weight[e], size_k, size_n, num_bits) + return output + + +def moe_sum(input: torch.Tensor, output: torch.Tensor): + ops.moe_sum(input, output) + + +def topk_softmax( + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + token_expert_indicies: torch.Tensor, + gating_output: float, +) -> None: + ops.topk_softmax(topk_weights, topk_ids, token_expert_indicies, gating_output) + + +__all__ = [ + "ScalarType", + "awq_marlin_moe_repack", + "fused_experts", + "fused_marlin_moe", + "fused_moe", + "fused_topk", + "gptq_marlin_moe_repack", + "grouped_topk", + "layers", + "moe_align_block_size", + "moe_sum", + "per_token_group_quant_fp8", + "scalar_types", + "topk_softmax", + "w8a8_block_fp8_matmul", +] diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/_moe_69e0359.abi3.so b/build/torch26-cxx11-cu126-aarch64-linux/moe/_moe_69e0359.abi3.so new file mode 100755 index 0000000000000000000000000000000000000000..0239658eb226f57499bfe6ffaec83c7d624315bd --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/_moe_69e0359.abi3.so @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb74055dc653477f6aad4757fd59b1e801d340d5101ab594201f9cc6c691b6ce +size 67586536 diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/_ops.py b/build/torch26-cxx11-cu126-aarch64-linux/moe/_ops.py new file mode 100644 index 0000000000000000000000000000000000000000..b33dfcc6afacf69f6989c184c5c3efabc03f18da --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/_ops.py @@ -0,0 +1,9 @@ +import torch +from . import _moe_69e0359 +ops = torch.ops._moe_69e0359 + +def add_op_namespace_prefix(op_name: str): + """ + Prefix op by namespace. + """ + return f"_moe_69e0359::{op_name}" \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..56c1a4e3af0b4a93fff71028d8e04bf73f0abb29 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..d3677bebb82a7f3f19344ef6471626493cf2c5bb --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=1792,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=1792,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..265768fb900ccfe9612b4a0d25973e6618f22a79 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=1792,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "5120": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "9216": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "13312": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "17408": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "25600": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "33792": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "41984": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "50176": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "58368": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..d3be23dfc903ba61d3d4d79c0230952b24d2ead0 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "5120": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "9216": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "13312": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "17408": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "25600": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "33792": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "41984": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "50176": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "58368": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=3072,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=3072,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..589f5d39f31418d5121e7cbb2e6f2894b0a7ed32 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=3072,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "5120": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "9216": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "13312": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "17408": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "25600": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "33792": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "41984": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "50176": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "58368": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=3072,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=3072,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..2c78bfaba7890772bf266721f5577202ea443882 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=3072,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "5120": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "9216": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "13312": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "17408": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "25600": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "33792": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "41984": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "50176": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "58368": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=3072,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=3072,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..4da841e74a79f9589fecac1fa557ea132d34805f --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=3072,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "5120": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "9216": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "13312": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "17408": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "25600": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "33792": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "41984": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "50176": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "58368": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=3584,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=3584,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..200356713c0d0a76e199671c7ec8f10d0e5ee0ac --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=3584,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "5120": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "9216": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "13312": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "17408": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "25600": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "33792": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "41984": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "50176": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "58368": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..e076615ee541a5043556f630ecf0946c4e2c1408 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "5120": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "9216": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "13312": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "17408": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "25600": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "33792": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "41984": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "50176": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "58368": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..ee896554b921040d7810bb6e9368cc200777951d --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "5120": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "9216": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "13312": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "17408": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "25600": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "33792": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "41984": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "50176": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "58368": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..05aed8b1c81492151d128ef251afc510d8cc8ed5 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "5120": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "9216": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "13312": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "17408": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "25600": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "33792": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "41984": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "50176": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "58368": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=1344,device_name=NVIDIA_A100-SXM4-40GB.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=1344,device_name=NVIDIA_A100-SXM4-40GB.json new file mode 100644 index 0000000000000000000000000000000000000000..9262a74a4a0e1e3789f260a3ef7f6cb9551f3f2b --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=1344,device_name=NVIDIA_A100-SXM4-40GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=1344,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=1344,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..d251f9b5accaec977fc87a0999cd56ee387fc650 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=1344,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=1344,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=1344,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..0ecf814a28a9441e89f892eb3d63dcf8dcb0dd97 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=1344,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=14336,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=14336,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..51ad5b299eb22465fa80530d12bdd5d7a03ce398 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=14336,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=14336,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=14336,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..ee5119182556cf49434c10e56cf04e3baeb26408 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=14336,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=1792,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=1792,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..68793c77b33c4f4b97d0a4b780fcbe8043c799de --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=1792,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "5120": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "9216": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "13312": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "17408": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "25600": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "33792": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "41984": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "50176": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "58368": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..612910720ed9439e56c4af4c03f30fee224fac80 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "5120": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "9216": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "13312": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "17408": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "25600": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "33792": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "41984": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "50176": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "58368": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=2688,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=2688,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..039a10ed127b77836a7f41c03513292613852b30 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=2688,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=2688,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=2688,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..3793fcafee60bc7e8f5f12d601cb3192abfa9ca8 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=2688,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=3072,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=3072,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..51d03d8607122d7b9bc20ba48d8432d62367fa00 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=3072,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=3072,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=3072,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..26f9abd6b789e9dd0f83ec7721fd1bae8aa76bec --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=3072,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "1536": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=3200,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=3200,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..cd0cdbea0c3372674cb610870dd0b30325864549 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=3200,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,130 @@ +{ + "3328": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "768": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "1792": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2560": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "2816": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "3584": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "3840": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1280": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "2304": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=3584,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=3584,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..64be6e6591422aa0f441c3747b6c49850929652e --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=3584,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..0a6a6a73fa45e270f01ba7ebdc6d9d55bf9daad3 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "5120": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "9216": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "13312": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "17408": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "25600": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "33792": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "41984": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "50176": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "58368": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=6400,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=6400,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..ba9041d008507e31ae4179ef2bc863a49c606582 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=6400,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,130 @@ +{ + "3840": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "1792": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "3584": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "2816": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1280": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "768": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "3328": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "2560": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "2304": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..7a7508aab04599cb06641c835d8b0a14f54d0716 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..dbf9a2dd6f048d8adee290961e2aea72035f7615 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..bbb2386046b1135a2cc7ab7cb26c1d0b039bcf3a --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "1536": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=800,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=800,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..57055453aa24c831dad9ac8e37fdab707c63ef91 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=800,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,130 @@ +{ + "2048": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "1792": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "3328": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "2560": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "768": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "2816": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "2304": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2 + }, + "1280": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "3840": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "3584": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=160,N=192,device_name=NVIDIA_A800-SXM4-80GB.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=160,N=192,device_name=NVIDIA_A800-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..0611620eb33626ac891454637723eabec9aa1183 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=160,N=192,device_name=NVIDIA_A800-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=1024,device_name=AMD_Instinct_MI325X,block_shape=[128,128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=1024,device_name=AMD_Instinct_MI325X,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..43c249d2530e1fafb46f6594b73e11f33fa8ffe7 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=1024,device_name=AMD_Instinct_MI325X,block_shape=[128,128].json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=1024,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=1024,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..43c249d2530e1fafb46f6594b73e11f33fa8ffe7 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=1024,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..4dd00d110e486c0a7670504ace778fd5995de727 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..48f9697af2639bd4afd45c768c7a8e9442924e2d --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..a8c05712ba5872db5414e571fef659572e59eb80 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..f1244c61efb019183e892f848700072c3be03ef9 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..2e692a1583a4ad95f020634d55f7f4e327be5da0 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128,128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..a2ee05da1d7c632aa6b1c3ccc8af2546f987ee93 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..63e118746fd86305fd03064637ad915b72c8abcd --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..e6769604ee6bcc75c41a5bea296cb55da3c755e8 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..e6769604ee6bcc75c41a5bea296cb55da3c755e8 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..fc573cd6e8561f56830c8f8ca3033f2d2f7bc1b3 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128,128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..c6d7e96c7f0aea61a2bd57723ecb91b1c1230895 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128,128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..6fcf408755f5d3e28ff636472f89e4fa52c8da22 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_L20,dtype=fp8_w8a8,block_shape=[128,128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_L20,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..c6eabea66a39aa86f5becc04d32ee01dbe627dce --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_L20,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=512,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=512,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..e6769604ee6bcc75c41a5bea296cb55da3c755e8 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=512,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=64,device_name=NVIDIA_A800-SXM4-80GB.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=64,device_name=NVIDIA_A800-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..21f60229ff875ca6dc0cd0a3cb4fa51a845bf881 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=64,device_name=NVIDIA_A800-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=60,N=1408,device_name=AMD_Instinct_MI300X.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=60,N=1408,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..d09508b31729ddbafc8b5e1dfa88343c61e83ca9 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=60,N=1408,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "8": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=60,N=176,device_name=AMD_Instinct_MI300X.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=60,N=176,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..746463af4d56315af3c31348c2253e43c926553e --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=60,N=176,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "3072": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=60,N=352,device_name=AMD_Instinct_MI300X.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=60,N=352,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..bbdb9ad096451d6c5b9d95fae4ed02ce37715002 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=60,N=352,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "2048": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "3072": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=60,N=704,device_name=AMD_Instinct_MI300X.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=60,N=704,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..43584b1eb6b65f02fc919490b0ff4b9569de58c4 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=60,N=704,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..8cc6c643f236d2f7f9ad29354d9e469d00b20d3f --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_A800-SXM4-80GB.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_A800-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..39a9912fa4bddc48de38d5a8b4fadd71326485d2 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_A800-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..05b54639d234e20096744915c932bb2d415641a5 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..d4c9ddd12972ac0b5fd2be11a9cd1075906e3978 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..c17a4ec346915af69f10f65410450cbd66614810 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H200.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H200.json new file mode 100644 index 0000000000000000000000000000000000000000..170ae7f3fff1d09961d368853dcfc2f24c1b878c --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H200.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=2560,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=2560,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..1d9d352edebc3f228353873677d2369e6a721158 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=2560,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=2560,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=2560,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..9ad5b316750051e4a5dc39b195f6a29eeaa882f0 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=2560,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=2560,device_name=NVIDIA_H200.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=2560,device_name=NVIDIA_H200.json new file mode 100644 index 0000000000000000000000000000000000000000..2883dfd11e7f3a7db429ae0aaeb98d554f271cdf --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=2560,device_name=NVIDIA_H200.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..8abfd84a776b7be2c0920ca8d3a69c06735a08ee --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..2fc18a5e43d298ddabc57dc9f357296890fb7e29 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..be8d4a7fd23d994ac838d45bf249e39764552ccf --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H200.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H200.json new file mode 100644 index 0000000000000000000000000000000000000000..71fdd88643c6ff515808fdc2d67e9fda3ac62ae9 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H200.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..b2799ed3a866e25b78d60d92910c000ebb21ff71 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_A800-SXM4-80GB.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_A800-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..c02de2f628b716456373978f600f9fc8c4b98d24 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_A800-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_GeForce_RTX_4090,dtype=fp8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_GeForce_RTX_4090,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..3e0bc75ff87c43ad3a0b3f132e5367a8ef39619a --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_GeForce_RTX_4090,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..9f7ed6726f44e60dea5af588e441b269b6c5cc3d --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..b8d3be2313fa14025d8aeb2fd11e0d1ee997ffa6 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..21b72557e365d01011f48a046915765e67dbd13f --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H200.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H200.json new file mode 100644 index 0000000000000000000000000000000000000000..eaf32f6d76c0a6aa50a5fbc2f055a5e7c9e10b66 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H200.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..b6f1d01f886524c58932ac28c31dcf6f10788e14 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..4bf775347eccb3e76e20ea0018d71343f9848313 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..f245285bd821717f2280fa89b5f0aa6bbce2262d --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI325X.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI325X.json new file mode 100644 index 0000000000000000000000000000000000000000..3918c93b160a39e55153cbf6ff634d59e724d021 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI325X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..3f3ccdafa88f3452a695efad4cb9622d6ae79e6a --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,138 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..841044a4fc6e2da81f18b2765fa30b9de7092b3d --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=NVIDIA_H200.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=NVIDIA_H200.json new file mode 100644 index 0000000000000000000000000000000000000000..59be497fc428755f010e490277f4b945d6ed0195 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=NVIDIA_H200.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..0e5fd1eec77d7e0aca0f1e3376b2c45217ce234a --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI300X.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..d6ad63509f1573d25fbdc370b93bc39a0d00b51a --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..16e0a91baf31d9a819f27eeea719e564f399379c --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI325X.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI325X.json new file mode 100644 index 0000000000000000000000000000000000000000..d766fc062ddcce10d38bdc68ab1f700d193b8ba7 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI325X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..8323f512db0158570427a7eb0a69bd889bc2d9ed --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI300X.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..1b46cb5716514f290c6b4bc3a18dba490c799272 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..6d5b1ae5b15fc333d3cdaa5dbcacab33267339cc --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI325X.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI325X.json new file mode 100644 index 0000000000000000000000000000000000000000..ffc1b23ea90ddefd24b4b98a222e67e84442cf9c --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI325X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_A100-SXM4-40GB.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_A100-SXM4-40GB.json new file mode 100644 index 0000000000000000000000000000000000000000..f4c0f8417b384870050a95e0cf57edbdf6352b23 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_A100-SXM4-40GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..5c8185cfdeec167ec4b88de51b4b395e28769cc5 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..97c9f4445b166657ad29f1db9fc8281f9c463ec4 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..e4110a5d2e70f4d2698d4c28b1668f8e4a48688a --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_H200.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_H200.json new file mode 100644 index 0000000000000000000000000000000000000000..0883ef40582ea4f1b08c629edcd6dd9beaaf2dad --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_H200.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..81bb765d30031ca3031636d7a442dc2c2292a6a9 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI300X.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..811c77ab4109319c9f079e414bd4f7d2cf4ac9f1 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..2758e48fc4065cd05ceb6fe22574acf2b3aae633 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI325X.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI325X.json new file mode 100644 index 0000000000000000000000000000000000000000..fc31215cbae8afeee27d923e176a25daf1f2b43d --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI325X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..0bb423b28f5ab3825929a4870b96393262a9dd9f --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..55571873395464a3b58f549523905f439a8f1716 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..26bcbf26970c7a77c99e2c8eacd83eefa86967bf --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..1a0aa331933291f0c32b2cf26672771b5c7b4196 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H200.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H200.json new file mode 100644 index 0000000000000000000000000000000000000000..9952be6ba4abe434e2924e9505e834de4ce0528e --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H200.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..379ca107a9469a5e5777e4e648f3f25ffc382a0a --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..5a3f415d54142383fa7677d8192f7a4c4df66b2e --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..6cb80f48329f5718e892fd6f78000df4003832f2 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI325X.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI325X.json new file mode 100644 index 0000000000000000000000000000000000000000..de9d0aba75a7ed45fdfd92c123736f1e981a06eb --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI325X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_A100-SXM4-40GB.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_A100-SXM4-40GB.json new file mode 100644 index 0000000000000000000000000000000000000000..b41f9d443e50678334f906b44fce6d018d69500e --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_A100-SXM4-40GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..edf2a38d12ad3f420f232d2cd61ab149ad138725 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_GeForce_RTX_4090,dtype=fp8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_GeForce_RTX_4090,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..32bbadbb9eae8d36e30895d863a8d10ff65167b2 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_GeForce_RTX_4090,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..673bae2ba8ef80ed4d4930739ca7daf0e8f28ee1 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..b2100cebb7f589747430be9ca8c8db368c152d78 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..e6f753cdba35b55e8d021597dc4911bab27caa8d --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H200.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H200.json new file mode 100644 index 0000000000000000000000000000000000000000..53f3394693f060daebff9960a95ed3753e8522c9 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H200.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_L40S.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_L40S.json new file mode 100644 index 0000000000000000000000000000000000000000..d720deb4bdd73d194b1023c99e190b8fcfecdaef --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_L40S.json @@ -0,0 +1,173 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 2 + }, + "2": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 2, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 7 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 128, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 1 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 1 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 2, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 2, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 2, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 2 + }, + "192": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_ctas": 1, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 128, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 8 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_ctas": 1, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_ctas": 1, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 2 + }, + "6144": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 2 + }, + "8192": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_ctas": 1, + "num_stages": 2 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..48bb5f2ccb8e3c623593538461d8d8ba71f93a69 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI300X.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..a64d06c6d1724f736c5081a8b947fa2e6127701d --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..2c49f359c22ac8d3b8fd3b00035b70472c34ed51 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI325X.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI325X.json new file mode 100644 index 0000000000000000000000000000000000000000..c7db6c0cbd3faf615074fc8055ff744e528dd204 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI325X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..dbc624731f5cb9afcdc9213183d00d1e5edd4a00 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..cc614e635ea57327c610ce79e99ae5339614f22e --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..32c0c9da471cbe479044095e0ed14a0f54b73620 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..4dd475c02a19b380efbe8f59945a0f2a41acda11 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H200.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H200.json new file mode 100644 index 0000000000000000000000000000000000000000..2ed15f30fe603c080a2e8a32b664400a50d2153e --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H200.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..bd2c6fbc1b941cebfe8e03f62e3044f5878db0ed --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..8d7b780271854c0fbae29e92cfd3cce508ca029e --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..7a07bbf4141952c4dc6199676b1ef97d1cfabece --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI325X.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI325X.json new file mode 100644 index 0000000000000000000000000000000000000000..3a3268cc17a84d88f0a67e20e266491ad45d1777 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI325X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..f578c8d0160ac3ef85b53c8539d3675455a97173 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..918f6839620cbab1f30b0f9383a9129c2cf2cf3d --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..e341a67917d5177bacb3f6767e7b6d92539826ad --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..eb817268d41208a661d2ab9526c5108d36d44085 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H200.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H200.json new file mode 100644 index 0000000000000000000000000000000000000000..0c7062aea6c4e920cdfdc9e13fc2c0fa38aceb0e --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H200.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..cd4fb8f11b935e401103aab5117a1d1be1799c6e --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI300X.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..cf66868e9d57a541fe10fa9fff682f7a7c6dac97 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..c27ca0a36594fd12a9995fffda60a9d5003425c0 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI325X.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI325X.json new file mode 100644 index 0000000000000000000000000000000000000000..da477b1fb15eb2064cbff78b5e47ae9c54dae7a5 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI325X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..34b916e574f88c65db1dac5889d74a990dc25e9b --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..96cbc111c7fffe8b0a67c22711f9a5c8f28d8b61 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=1536,K=1536,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=1536,K=1536,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..6496a38fba8ae09b3025a75f357815b9d6a5e3f4 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=1536,K=1536,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=1536,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=1536,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..3618053b65831b95c4bb0f20ef3b9aa816b2d637 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=1536,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=1536,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=1536,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..46a982f5ee9a4bd67ce244b101c576efeeb53b78 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=1536,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=2048,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=2048,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..035ec027fa56622196b24a03a5042ce010deaebf --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=2048,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=2048,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=2048,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..8b49f2781cb54d19a2789767ebb7e8c3fb55b981 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=2048,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=2304,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=2304,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..851bc9f9f0b50b41451b929eaa518869b6a05412 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=2304,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=2304,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=2304,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..d1227c2157990216d2ca51c69ad0944017f53b6a --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=2304,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=24576,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=24576,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..77ba0d7477bdbcb036a43263e7aaa6b6913f8f4e --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=24576,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=24576,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=24576,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..1c61451fb34e52deec827f8f63c80fb15830c202 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=24576,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=24576,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=24576,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..63e661c80de6a7b1422f7a994a2ee7a4b724911c --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=24576,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=256,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=256,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..cf354037903c0d1fcd077c4647aabce026a723fb --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=256,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=3072,K=1536,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=3072,K=1536,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..0a5d7bfdba4852da9ed08d1bc27cd7d521d09965 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=3072,K=1536,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=3072,K=1536,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=3072,K=1536,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..eccb86a76df0d7302b760ab6d83a8ceb9fa9d0d9 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=3072,K=1536,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=3072,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=3072,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..cb91a279d423d0ca25197e0edd5e8c2f4da58720 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=3072,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=3072,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=3072,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..88af48431d8b8791af8df03429704606b670f1f7 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=3072,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=3072,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=3072,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..dd069726d7ed4dcbb449af243f4f4af21815f854 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=3072,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=32768,K=512,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=32768,K=512,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..7febe3d272b4bb76500f7c6b523396129fd53680 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=32768,K=512,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=32768,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=32768,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..56b939e52fac3ed53a4e0ba640c40010cb3af30a --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=32768,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=32768,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=32768,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..63d9a0bf5d79ddaaad547d44338ad4b959ad72b1 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=32768,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=36864,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=36864,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..7fa398c15a2a535401709b0f25e20f6e4b23e58e --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=36864,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=36864,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=36864,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..f15d8f64c7090bd71d0091a524c65d7818fec38e --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=36864,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=4096,K=512,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=4096,K=512,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..9d7658bfc41b2c8fd4daf3fbdf62d15936d3d546 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=4096,K=512,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=4096,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=4096,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..cd3e07804fdec10c2cfb291c1ede3ba67b753f9c --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=4096,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=4096,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=4096,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..9d5a329d7466a37c0ca68a65a089fbb99f9327a9 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=4096,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=4608,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=4608,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..03dba5ad15ba5f7f49100a5c78e8685e64334b2a --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=4608,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=4608,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=4608,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..96e1594a3eabbaedc792b84b07f05ae8752b7251 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=4608,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=4608,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=4608,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..5ffd367df833d773355590220598a3c7eceba4e0 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=4608,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=512,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=512,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..9a5ff48b8942957dde9b862aed848390dd267948 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=512,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=512,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=512,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..eabc423949a24c2a1fb2368a73e5249caf8d07df --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=512,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=576,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=576,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..386928de139ce718f28222b9c1a6555df3958491 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=576,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 5 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=576,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=576,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..51e237b91b8e775a36bcf783c078c2c1cecbcbd2 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=576,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=576,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=576,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..6280219c9ee7d26f7e2fd3625dc92d847ddc7982 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=576,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=1024,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=1024,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..40c01c0b92b4b26fe480879dda33f18c5eb59a6d --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=1024,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=1024,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=1024,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..c6fd3659799bc31e17f3577e7f0e8d7268faf1fb --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=1024,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=1152,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=1152,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..160f12ed3f95a6967439ff53bc3e3a2cdc97c700 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=1152,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=1152,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=1152,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..e5c4a1d2c94e5c7864f462e083ea5f530b8efe3f --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=1152,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=128,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=128,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..2bf5eb27e38208871d50348b170c8c74b80fc519 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=128,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "96": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=16384,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=16384,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..9c908e80406587da4d246ce4e3a8a98a14c875b1 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=16384,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=16384,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=16384,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..0a1e14cffbb2a894a701352193947d272427db0d --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=16384,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=16384,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=16384,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..15b1c93f60fc5068ba11b82b6d5924dd2024a824 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=16384,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=18432,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=18432,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..f78e7060e6840ff721d306db556636b0bbc8d9b3 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=18432,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=18432,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=18432,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..8ff12e64c172f5a5d0fbdf900728fe60b33877e2 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=18432,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=18432,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=18432,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..4532f93681e2be175b1bf94f81bfde711821cd60 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=18432,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=2048,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=2048,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..1d3ce5c94c2d9a4a1637204efb3b14f7a5579bdb --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=2048,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=2048,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=2048,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..ca7f32b9552b479dc05495792b7e426db5eb1b56 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=2048,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=2048,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=2048,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..5acea242cc0ad094cba8ee5f568ff88afb1b41ae --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=2048,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=2304,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=2304,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..3ab5796ee15b6ec8d4ab1f4ab5a594fecb30e4b4 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=2304,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=2304,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=2304,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..58cdd93e90b8c29bc7a211861711565dbeeb529a --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=2304,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=2304,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=2304,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..b72e0371d1421a1decc9d57860f83eea8f790942 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=2304,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=256,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=256,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..3cb7eaa07c745fd3aa2b3242780a7061bedac1de --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=256,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..293adce387e066fce75b6e606d4b8b6a5aa10bdb --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + } +} diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/fp8.py b/build/torch26-cxx11-cu126-aarch64-linux/moe/fp8.py new file mode 100644 index 0000000000000000000000000000000000000000..23bd7d6703104b0020671cc6ba6f78a6df37e4bf --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/fp8.py @@ -0,0 +1,63 @@ +from typing import Optional, Tuple, Union + +import torch + +from ._ops import ops + + +def is_hip() -> bool: + return torch.version.hip is not None + + +def scaled_fp8_quant( + input: torch.Tensor, + scale: Optional[torch.Tensor] = None, + num_token_padding: Optional[int] = None, + scale_ub: Optional[torch.Tensor] = None, + use_per_token_if_dynamic: bool = False, +) -> Tuple[torch.Tensor, torch.Tensor]: + """ + Quantize input tensor to FP8 and return quantized tensor and scale. + + This function supports both static and dynamic quantization: If you + provide the scale, it will use static scaling and if you omit it, + the scale will be determined dynamically. The function also allows + optional padding of the output tensors for downstream kernels that + will benefit from padding. + + Args: + input: The input tensor to be quantized to FP8 + scale: Optional scaling factor for the FP8 quantization + scale_ub: Optional upper bound for scaling factor in dynamic + per token case + num_token_padding: If specified, pad the first dimension + of the output to at least this value. + use_per_token_if_dynamic: Whether to do per_tensor or per_token + in the dynamic quantization case. + + Returns: + Tuple[torch.Tensor, torch.Tensor]: The output tensor in FP8 and + scaling factor. + """ + # This code assumes batch_dim and num_tokens are flattened + assert input.ndim == 2 + shape: Union[Tuple[int, int], torch.Size] = input.shape + # For rocm, the output fp8 dtype is torch.float_e3m3fnuz + out_dtype: torch.dtype = torch.float8_e4m3fnuz if is_hip() else torch.float8_e4m3fn + if num_token_padding: + shape = (max(num_token_padding, input.shape[0]), shape[1]) + output = torch.empty(shape, device=input.device, dtype=out_dtype) + + if scale is None: + if use_per_token_if_dynamic: + scale = torch.empty((shape[0], 1), device=input.device, dtype=torch.float32) + ops.dynamic_per_token_scaled_fp8_quant(output, input, scale, scale_ub) + else: + scale = torch.zeros(1, device=input.device, dtype=torch.float32) + ops.dynamic_scaled_fp8_quant(output, input, scale) + else: + # num_token_padding not implemented for this case + assert scale.numel() == 1 or num_token_padding is None + ops.static_scaled_fp8_quant(output, input, scale) + + return output, scale diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/fp8_utils.py b/build/torch26-cxx11-cu126-aarch64-linux/moe/fp8_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..acb4f3e3bb1a34f209fdac9ecca8c123aaf67f12 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/fp8_utils.py @@ -0,0 +1,391 @@ +import functools +import json +import logging +import os +from typing import Any, Dict, List, Optional, Tuple + +import torch +import triton +import triton.language as tl + +from .platforms import current_platform + +logger = logging.getLogger(__name__) + + +@triton.jit +def _per_token_group_quant_fp8( + # Pointers to inputs and output + y_ptr, + y_q_ptr, + y_s_ptr, + group_size, + # Avoid to divide zero + eps, + # Information for float8 + fp8_min, + fp8_max, + # Meta-parameters + BLOCK: tl.constexpr, +): + """A Triton-accelerated function to perform per-token-group + quantization on a tensor. + This function converts the tensor values into float8 values. + """ + # Map the program id to the row of X and Y it should compute. + g_id = tl.program_id(0) + y_ptr += g_id * group_size + y_q_ptr += g_id * group_size + y_s_ptr += g_id + + cols = tl.arange(0, BLOCK) # N <= BLOCK + mask = cols < group_size + + y = tl.load(y_ptr + cols, mask=mask, other=0.0).to(tl.float32) + # Quant + _absmax = tl.maximum(tl.max(tl.abs(y)), eps) + y_s = _absmax / fp8_max + y_q = tl.clamp(y / y_s, fp8_min, fp8_max).to(y_q_ptr.dtype.element_ty) + + tl.store(y_q_ptr + cols, y_q, mask=mask) + tl.store(y_s_ptr, y_s) + + +@triton.jit +def _per_token_group_quant_fp8_colmajor( + # Pointers to inputs and output + y_ptr, + y_q_ptr, + y_s_ptr, + group_size, + # Num columns of y + y_num_columns, + # Stride from one column to the next of y_s + y_s_col_stride, + # Avoid to divide zero + eps, + # Information for float8 + fp8_min, + fp8_max, + # Meta-parameters + BLOCK: tl.constexpr, +): + """A Triton-accelerated function to perform per-token-group + quantization on a tensor. + This function converts the tensor values into float8 values. + """ + # Map the program id to the row of X and Y it should compute. + g_id = tl.program_id(0) + y_ptr += g_id * group_size + y_q_ptr += g_id * group_size + + # Convert g_id the flattened block coordinate to 2D so we can index + # into the output y_scales matrix + blocks_per_row = y_num_columns // group_size + scale_col = g_id % blocks_per_row + scale_row = g_id // blocks_per_row + y_s_ptr += scale_col * y_s_col_stride + scale_row + + cols = tl.arange(0, BLOCK) # group_size <= BLOCK + mask = cols < group_size + + y = tl.load(y_ptr + cols, mask=mask, other=0.0).to(tl.float32) + # Quant + _absmax = tl.maximum(tl.max(tl.abs(y)), eps) + y_s = _absmax / fp8_max + y_q = tl.clamp(y / y_s, fp8_min, fp8_max).to(y_q_ptr.dtype.element_ty) + + tl.store(y_q_ptr + cols, y_q, mask=mask) + tl.store(y_s_ptr, y_s) + + +def per_token_group_quant_fp8( + x: torch.Tensor, + group_size: int, + eps: float = 1e-10, + dtype: Optional[torch.dtype] = None, + column_major_scales: bool = False, +) -> Tuple[torch.Tensor, torch.Tensor]: + """Function to perform per-token-group quantization on an input tensor `x`. + It converts the tensor values into signed float8 values and returns the + quantized tensor along with the scaling factor used for quantization. + Args: + x: The input tensor with ndim >= 2. + group_size: The group size used for quantization. + eps: The minimum to avoid dividing zero. + dtype: The dype of output tensor. Note that only `torch.float8_e4m3fn` + is supported for now. + Returns: + Tuple[torch.Tensor, torch.Tensor]: The quantized tensor and the + scaling factor for quantization. + """ + if dtype is None: + dtype = ( + torch.float8_e4m3fnuz if current_platform.is_rocm() else torch.float8_e4m3fn + ) + assert x.shape[-1] % group_size == 0, ( + f"the last dimension of `x` {x.shape[-1]} must be divisible " + f"by `group_size` {group_size}" + ) + assert x.is_contiguous(), "`x` must be contiguous" + + finfo = torch.finfo(dtype) + fp8_min = finfo.min + fp8_max = finfo.max + + x_q = torch.empty_like(x, device=x.device, dtype=dtype) + M = x.numel() // group_size + N = group_size + if column_major_scales: + shape = (x.shape[-1] // group_size,) + x.shape[:-1] + x_s = torch.empty(shape, device=x.device, dtype=torch.float32).permute(-1, -2) + else: + shape = x.shape[:-1] + (x.shape[-1] // group_size,) + x_s = torch.empty(shape, device=x.device, dtype=torch.float32) + + BLOCK = triton.next_power_of_2(N) + # heuristics for number of warps + num_warps = min(max(BLOCK // 256, 1), 8) + num_stages = 1 + if column_major_scales: + _per_token_group_quant_fp8_colmajor[(M,)]( + x, + x_q, + x_s, + group_size, + x.shape[1], + x_s.stride(1), + eps, + fp8_min=fp8_min, + fp8_max=fp8_max, + BLOCK=BLOCK, + num_warps=num_warps, + num_stages=num_stages, + ) + else: + _per_token_group_quant_fp8[(M,)]( + x, + x_q, + x_s, + group_size, + eps, + fp8_min=fp8_min, + fp8_max=fp8_max, + BLOCK=BLOCK, + num_warps=num_warps, + num_stages=num_stages, + ) + + return x_q, x_s + + +@triton.jit +def _w8a8_block_fp8_matmul( + # Pointers to inputs and output + A, + B, + C, + As, + Bs, + # Shape for matmul + M, + N, + K, + # Block size for block-wise quantization + group_n, + group_k, + # Stride for inputs and output + stride_am, + stride_ak, + stride_bk, + stride_bn, + stride_cm, + stride_cn, + stride_As_m, + stride_As_k, + stride_Bs_k, + stride_Bs_n, + # Meta-parameters + BLOCK_SIZE_M: tl.constexpr, + BLOCK_SIZE_N: tl.constexpr, + BLOCK_SIZE_K: tl.constexpr, + GROUP_SIZE_M: tl.constexpr, +): + """Triton-accelerated function used to perform linear operations (dot + product) on input tensors `A` and `B` with block-wise quantization, and + store the result in output tensor `C`. + """ + + pid = tl.program_id(axis=0) + num_pid_m = tl.cdiv(M, BLOCK_SIZE_M) + num_pid_n = tl.cdiv(N, BLOCK_SIZE_N) + num_pid_in_group = GROUP_SIZE_M * num_pid_n + group_id = pid // num_pid_in_group + first_pid_m = group_id * GROUP_SIZE_M + group_size_m = min(num_pid_m - first_pid_m, GROUP_SIZE_M) + pid_m = first_pid_m + (pid % group_size_m) + pid_n = (pid % num_pid_in_group) // group_size_m + + offs_am = (pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)) % M + offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)) % N + offs_k = tl.arange(0, BLOCK_SIZE_K) + a_ptrs = A + (offs_am[:, None] * stride_am + offs_k[None, :] * stride_ak) + b_ptrs = B + (offs_k[:, None] * stride_bk + offs_bn[None, :] * stride_bn) + + As_ptrs = As + offs_am * stride_As_m + offs_bsn = offs_bn // group_n + Bs_ptrs = Bs + offs_bsn * stride_Bs_n + + accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32) + for k in range(0, tl.cdiv(K, BLOCK_SIZE_K)): + a = tl.load(a_ptrs, mask=offs_k[None, :] < K - k * BLOCK_SIZE_K, other=0.0) + b = tl.load(b_ptrs, mask=offs_k[:, None] < K - k * BLOCK_SIZE_K, other=0.0) + + k_start = k * BLOCK_SIZE_K + offs_ks = k_start // group_k + a_s = tl.load(As_ptrs + offs_ks * stride_As_k) + b_s = tl.load(Bs_ptrs + offs_ks * stride_Bs_k) + + accumulator += tl.dot(a, b) * a_s[:, None] * b_s[None, :] + a_ptrs += BLOCK_SIZE_K * stride_ak + b_ptrs += BLOCK_SIZE_K * stride_bk + + if C.dtype.element_ty == tl.bfloat16: + c = accumulator.to(tl.bfloat16) + elif C.dtype.element_ty == tl.float16: + c = accumulator.to(tl.float16) + else: + c = accumulator.to(tl.float32) + + offs_cm = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M) + offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N) + c_ptrs = C + stride_cm * offs_cm[:, None] + stride_cn * offs_cn[None, :] + c_mask = (offs_cm[:, None] < M) & (offs_cn[None, :] < N) + tl.store(c_ptrs, c, mask=c_mask) + + +@functools.lru_cache +def get_w8a8_block_fp8_configs( + N: int, K: int, block_n: int, block_k: int +) -> Optional[Dict[int, Any]]: + """ + Return optimized configurations for the w8a8 block fp8 kernel. + The return value will be a dictionary that maps an irregular grid of + batch sizes to configurations of the w8a8 block fp8 kernel. To evaluate the + kernel on a given batch size bs, the closest batch size in the grid should + be picked and the associated configuration chosen to invoke the kernel. + """ + + # First look up if an optimized configuration is available in the configs + # directory + device_name = current_platform.get_device_name().replace(" ", "_") + json_file_name = f"N={N},K={K},device_name={device_name},dtype=fp8_w8a8,block_shape=[{block_n}, {block_k}].json" # noqa: E501 + + config_file_path = os.path.join( + os.path.dirname(os.path.realpath(__file__)), "configs", json_file_name + ) + if os.path.exists(config_file_path): + with open(config_file_path) as f: + logger.info( + "Using configuration from %s for W8A8 Block FP8 kernel.", + config_file_path, + ) + # If a configuration has been found, return it + return {int(key): val for key, val in json.load(f).items()} + + # If no optimized configuration is available, we will use the default + # configuration + logger.warning( + "Using default W8A8 Block FP8 kernel config. Performance might " + "be sub-optimal! Config file not found at %s", + config_file_path, + ) + return None + + +def w8a8_block_fp8_matmul( + A: torch.Tensor, + B: torch.Tensor, + As: torch.Tensor, + Bs: torch.Tensor, + block_size: List[int], + output_dtype: torch.dtype = torch.float16, +) -> torch.Tensor: + """This function performs matrix multiplication with block-wise + quantization. + It takes two input tensors `A` and `B` with scales `As` and `Bs`. + The output is returned in the specified `output_dtype`. + Args: + A: The input tensor, e.g., activation. + B: The input tensor, e.g., weight. + As: The per-token-group quantization scale for `A`. + Bs: The per-block quantization scale for `B`. + block_size: The block size for per-block quantization. It should + be 2-dim, e.g., [128, 128]. + output_dytpe: The dtype of the returned tensor. + Returns: + torch.Tensor: The result of matmul. + """ + assert len(block_size) == 2 + block_n, block_k = block_size[0], block_size[1] + + assert A.shape[-1] == B.shape[-1] + assert A.shape[:-1] == As.shape[:-1] and A.is_contiguous() + assert triton.cdiv(A.shape[-1], block_k) == As.shape[-1] + M = A.numel() // A.shape[-1] + + assert B.ndim == 2 and B.is_contiguous() and Bs.ndim == 2 + N, K = B.shape + assert triton.cdiv(N, block_n) == Bs.shape[0] + assert triton.cdiv(K, block_k) == Bs.shape[1] + + C_shape = A.shape[:-1] + (N,) + C = A.new_empty(C_shape, dtype=output_dtype) + + configs = get_w8a8_block_fp8_configs(N, K, block_size[0], block_size[1]) + if configs: + # Get the optimal config if there is one + config = configs[min(configs.keys(), key=lambda x: abs(x - M))] + else: + # Default config + # Block-wise quant: BLOCK_SIZE_N must be divisible by block_size[0] + # BLOCK_SIZE_K must be divisible by block_size[1] + config = { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": block_size[0], + "BLOCK_SIZE_K": block_size[1], + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2, + } + + def grid(META): + return ( + triton.cdiv(M, META["BLOCK_SIZE_M"]) * triton.cdiv(N, META["BLOCK_SIZE_N"]), + ) + + _w8a8_block_fp8_matmul[grid]( + A, + B, + C, + As, + Bs, + M, + N, + K, + block_n, + block_k, + A.stride(-2), + A.stride(-1), + B.stride(1), + B.stride(0), + C.stride(-2), + C.stride(-1), + As.stride(-2), + As.stride(-1), + Bs.stride(1), + Bs.stride(0), + **config, + ) + + return C diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/fused_marlin_moe.py b/build/torch26-cxx11-cu126-aarch64-linux/moe/fused_marlin_moe.py new file mode 100644 index 0000000000000000000000000000000000000000..b3e0a5c24599730faf973fad3cf3fb6031a30522 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/fused_marlin_moe.py @@ -0,0 +1,413 @@ +"""Fused MoE utilities for GPTQ.""" + +import functools +from typing import TYPE_CHECKING, Optional + +import torch + +from ._ops import add_op_namespace_prefix, ops +from .fused_moe import fused_topk, moe_align_block_size, try_get_optimal_moe_config +from .scalar_type import ScalarType, scalar_types + +# neuron has torch version that doesn't even have impl_abstract +if TYPE_CHECKING: + + def register_fake(fn): + return lambda name: fn + +else: + try: + from torch.library import register_fake + except ImportError: + from torch.library import impl_abstract as register_fake + + +def get_scalar_type(num_bits: int, has_zp: bool): + if has_zp: + assert num_bits == 4 + return scalar_types.uint4 + else: + return scalar_types.uint4b8 if num_bits == 4 else scalar_types.uint8b128 + + +def single_marlin_moe( + hidden_states: torch.Tensor, + w: torch.Tensor, + scales: torch.Tensor, + gating_output: torch.Tensor, + topk: int, + renormalize: bool, + g_idx: Optional[torch.Tensor] = None, + sort_indices: Optional[torch.Tensor] = None, + w_zeros: Optional[torch.Tensor] = None, + num_bits: int = 8, + is_k_full: bool = True, +) -> torch.Tensor: + """ + This function computes the multiplication of hidden_states with expert + weights used in Marlin MoE, using weights w and top-k gating mechanism. + Its purpose is testing and debugging the fused MoE kernel. + + Parameters: + - hidden_states (torch.Tensor): The input tensor to the Marlin Mul. + - w (torch.Tensor): The set of expert weights. + - scales (torch.Tensor): The quantization scales. + - gating_output (torch.Tensor): The output of the gating operation + (before softmax). + - g_idx (Optional[torch.Tensor]): Optional act_order indices. + - sort_indices (Optional[torch.Tensor]): Optional act_order input + permutation. + - topk (int): The number of top-k experts to select. + - renormalize (bool): If True, renormalize the top-k weights to sum to 1. + - w_zeros (Optional[torch.Tensor]): Optional zero points to be used for w. + - num_bits (bool): The number of bits in expert weights quantization. + + Returns: + - torch.Tensor: The output tensor after applying the MoE layer. + """ + # Check constraints. + assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch" + assert hidden_states.shape[1] == w.shape[1] * 16, "Hidden size mismatch" + assert gating_output.shape[1] == w.shape[0], "Number of experts mismatch" + assert hidden_states.is_contiguous(), "Hidden_states must be contiguous" + assert w.is_contiguous(), "Expert weights must be contiguous" + assert hidden_states.dtype == torch.float16 + assert num_bits in [4, 8] + + M, K = hidden_states.shape + E = w.shape[0] + N = w.shape[2] // (num_bits // 2) + + topk_weights, topk_ids = fused_topk(hidden_states, gating_output, topk, renormalize) + + # This might not be an optimal config for a single MMM + get_config_func = functools.partial( + try_get_optimal_moe_config, + w.shape, + w.shape, + topk_ids.shape[1], + None, + is_marlin=True, + ) + config = get_config_func(M) + + block_size_m = config["BLOCK_SIZE_M"] + + sorted_token_ids, _, _ = moe_align_block_size(topk_ids, block_size_m, E) + + max_workspace_size = (N // 64) * 16 + workspace = torch.zeros( + max_workspace_size, + dtype=torch.int, + device=hidden_states.device, + requires_grad=False, + ) + + has_zero_point = w_zeros is not None + if w_zeros is None: + w_zeros = torch.empty( + (0, 0), + dtype=hidden_states.dtype, + device=hidden_states.device, + requires_grad=False, + ) + + if g_idx is None: + g_idx = torch.empty( + (0, 0), dtype=torch.int32, device=hidden_states.device, requires_grad=False + ) + + if sort_indices is None: + sort_indices = torch.empty( + (0), dtype=torch.int32, device=hidden_states.device, requires_grad=False + ) + + scalar_type = get_scalar_type(num_bits, has_zero_point) + + intermediate_cache = ops.marlin_gemm_moe( + hidden_states, + w, + sorted_token_ids, + topk_weights, + topk_ids, + scales, + w_zeros, + g_idx, + sort_indices, + workspace, + scalar_type.id, + M, + N, + K, + is_k_full, + E, + topk, + block_size_m, + True, + False, + ) + + return torch.sum(intermediate_cache.view(*intermediate_cache.shape), dim=1) + + +if hasattr(ops, "single_marlin_gemm_moe"): + + @register_fake(add_op_namespace_prefix("single_marlin_gemm_moe")) + def single_marlin_moe_fake( + hidden_states: torch.Tensor, + w: torch.Tensor, + scales: torch.Tensor, + gating_output: torch.Tensor, + topk: int, + renormalize: bool, + g_idx: Optional[torch.Tensor] = None, + sort_indices: Optional[torch.Tensor] = None, + w_zeros: Optional[torch.Tensor] = None, + num_bits: int = 8, + is_k_full: bool = True, + ) -> torch.Tensor: + return torch.empty_like(hidden_states) + + +def fused_marlin_moe( + hidden_states: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + w1_scale: torch.Tensor, + w2_scale: torch.Tensor, + gating_output: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + g_idx1: Optional[torch.Tensor] = None, + g_idx2: Optional[torch.Tensor] = None, + sort_indices1: Optional[torch.Tensor] = None, + sort_indices2: Optional[torch.Tensor] = None, + w1_zeros: Optional[torch.Tensor] = None, + w2_zeros: Optional[torch.Tensor] = None, + num_bits: int = 8, + is_k_full: bool = True, +) -> torch.Tensor: + """ + This function computes a Mixture of Experts (MoE) layer using two sets of + weights, w1 and w2, and top-k gating mechanism. + + Parameters: + - hidden_states (torch.Tensor): The input tensor to the MoE layer. + - w1 (torch.Tensor): The first set of expert weights. + - w2 (torch.Tensor): The second set of expert weights. + - w1_scale (torch.Tensor): Scale to be used for w1. + - w2_scale (torch.Tensor): Scale to be used for w2. + - gating_output (torch.Tensor): The output of the gating operation + (before softmax). + - g_idx1 (Optional[torch.Tensor]): The first set of act_order indices. + - g_idx2 (Optional[torch.Tensor]): The second set of act_order indices. + - sort_indices1 (Optional[torch.Tensor]): The first act_order input + permutation. + - sort_indices2 (Optional[torch.Tensor]): The second act_order input + permutation. + - topk_weights (torch.Tensor): Top-k weights. + - topk_ids (torch.Tensor): Indices of topk-k elements. + - w1_zeros (Optional[torch.Tensor]): Optional zero points to be used for w1. + - w2_zeros (Optional[torch.Tensor]): Optional zero points to be used for w2. + - num_bits (bool): The number of bits in expert weights quantization. + + Returns: + - torch.Tensor: The output tensor after applying the MoE layer. + """ + # Check constraints. + assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch" + assert hidden_states.shape[1] == w1.shape[1] * 16, "Hidden size mismatch w1" + assert hidden_states.shape[1] == w2.shape[2] // ( + num_bits // 2 + ), "Hidden size mismatch w2" + assert gating_output.shape[1] == w1.shape[0], "Number of experts mismatch" + assert hidden_states.is_contiguous(), "Hidden_states must be contiguous" + assert w1.is_contiguous(), "Expert weights1 must be contiguous" + assert w2.is_contiguous(), "Expert weights2 must be contiguous" + assert hidden_states.dtype == torch.float16 + assert num_bits in [4, 8] + + has_no_act_order = ( + g_idx1 is None + and g_idx2 is None + and sort_indices1 is None + and sort_indices2 is None + ) + has_all_act_order = ( + g_idx1 is not None + and g_idx2 is not None + and sort_indices1 is not None + and sort_indices2 is not None + ) + assert has_no_act_order or has_all_act_order, ( + "g_idx and sorted_indices " "must be all not None or must be all None" + ) + + has_no_zp = w1_zeros is None and w2_zeros is None + has_all_zp = w1_zeros is not None and w2_zeros is not None + assert has_no_zp or has_all_zp, ( + "zero points must be both not None or " "must be both None" + ) + + M, K = hidden_states.shape + E = w1.shape[0] + N = w2.shape[1] * 16 + topk = topk_ids.shape[1] + + get_config_func = functools.partial( + try_get_optimal_moe_config, + w1.shape, + w2.shape, + topk_ids.shape[1], + None, + is_marlin=True, + ) + config = get_config_func(M) + + block_size_m = config["BLOCK_SIZE_M"] + + sorted_token_ids, _, _ = moe_align_block_size(topk_ids, block_size_m, E) + + max_workspace_size = (max(2 * N, K) // 64) * 16 + workspace = torch.zeros( + max_workspace_size, dtype=torch.int, device="cuda", requires_grad=False + ) + + if has_no_zp: + w1_zeros = torch.empty( + (0, 0), + dtype=hidden_states.dtype, + device=hidden_states.device, + requires_grad=False, + ) + w2_zeros = torch.empty( + (0, 0), + dtype=hidden_states.dtype, + device=hidden_states.device, + requires_grad=False, + ) + + if has_no_act_order: + g_idx1 = torch.empty( + (0, 0), dtype=torch.int32, device=hidden_states.device, requires_grad=False + ) + g_idx2 = torch.empty( + (0, 0), dtype=torch.int32, device=hidden_states.device, requires_grad=False + ) + sort_indices1 = torch.empty( + (0), dtype=torch.int32, device=hidden_states.device, requires_grad=False + ) + sort_indices2 = torch.empty( + (0, 0), dtype=torch.int32, device=hidden_states.device, requires_grad=False + ) + + scalar_type1 = get_scalar_type(num_bits, has_all_zp) + scalar_type2 = get_scalar_type(num_bits, has_all_zp) + + intermediate_cache2 = torch.empty( + (M * topk_ids.shape[1], N), + device=hidden_states.device, + dtype=hidden_states.dtype, + ) + + intermediate_cache1 = ops.marlin_gemm_moe( + hidden_states, + w1, + sorted_token_ids, + topk_weights, + topk_ids, + w1_scale, + w1_zeros, + g_idx1, + sort_indices1, + workspace, + scalar_type1.id, + M, + 2 * N, + K, + is_k_full, + E, + topk, + block_size_m, + True, + False, + ) + + ops.silu_and_mul(intermediate_cache2, intermediate_cache1.view(-1, 2 * N)) + + intermediate_cache3 = ops.marlin_gemm_moe( + intermediate_cache2, + w2, + sorted_token_ids, + topk_weights, + topk_ids, + w2_scale, + w2_zeros, + g_idx2, + sort_indices2, + workspace, + scalar_type2.id, + M, + K, + N, + is_k_full, + E, + topk, + block_size_m, + False, + True, + ) + + return torch.sum(intermediate_cache3.view(*intermediate_cache3.shape), dim=1) + + +if hasattr(ops, "fused_marlin_moe"): + + @register_fake(add_op_namespace_prefix("fused_marlin_moe")) + def fused_marlin_moe_fake( + hidden_states: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + w1_scale: torch.Tensor, + w2_scale: torch.Tensor, + gating_output: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + g_idx1: Optional[torch.Tensor] = None, + g_idx2: Optional[torch.Tensor] = None, + sort_indices1: Optional[torch.Tensor] = None, + sort_indices2: Optional[torch.Tensor] = None, + w1_zeros: Optional[torch.Tensor] = None, + w2_zeros: Optional[torch.Tensor] = None, + num_bits: int = 8, + is_k_full: bool = True, + ) -> torch.Tensor: + return torch.empty_like(hidden_states) + + +if hasattr(ops, "marlin_gemm_moe"): + + @register_fake(add_op_namespace_prefix("marlin_gemm_moe")) + def marlin_gemm_moe_fake( + a: torch.Tensor, + b_q_weights: torch.Tensor, + sorted_ids: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + b_scales: torch.Tensor, + b_zero_points: torch.Tensor, + g_idx: torch.Tensor, + perm: torch.Tensor, + workspace: torch.Tensor, + b_q_type: ScalarType, + size_m: torch.SymInt, + size_n: torch.SymInt, + size_k: torch.SymInt, + is_k_full: bool, + num_experts: int, + topk: int, + moe_block_size: int, + replicate_input: bool, + apply_weights: bool, + ) -> torch.Tensor: + return torch.empty((size_m, topk, size_n), dtype=a.dtype, device=a.device) diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/fused_moe.py b/build/torch26-cxx11-cu126-aarch64-linux/moe/fused_moe.py new file mode 100644 index 0000000000000000000000000000000000000000..f8f0586c2a58d1fb68f7d2eaed47a48f165d1e8d --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/fused_moe.py @@ -0,0 +1,1574 @@ +# SPDX-License-Identifier: Apache-2.0 +"""Fused MoE kernel.""" + +import functools +import json +import logging +import os +from typing import Any, Callable, Dict, List, Optional, Tuple + +import torch +import triton +import triton.language as tl + +from ._ops import ops +from .fp8 import scaled_fp8_quant +from .fp8_utils import per_token_group_quant_fp8 +from .moe_align_block_size import moe_align_block_size +from .platforms import current_platform +from .utils.custom_ops import direct_register_custom_op + +logger = logging.getLogger(__name__) + + +VLLM_FUSED_MOE_CHUNK_SIZE = int(os.getenv("VLLM_FUSED_MOE_CHUNK_SIZE", "32768")) +VLLM_ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON: bool = bool( + int(os.getenv("VLLM_ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON", "0")) +) + + +def cdiv(a: int, b: int) -> int: + """Ceiling division.""" + return -(a // -b) + + +def _fp8_quantize( + A: torch.Tensor, + A_scale: Optional[torch.Tensor], + block_shape: Optional[List[int]], +) -> Tuple[torch.Tensor, torch.Tensor]: + """ + Perform fp8 quantization on the inputs. If a block_shape + is provided, the output will be blocked. + """ + if block_shape is None: + A, A_scale = scaled_fp8_quant(A, A_scale) + else: + assert len(block_shape) == 2 + _, block_k = block_shape[0], block_shape[1] + A, A_scale = per_token_group_quant_fp8(A, block_k) + assert cdiv(A.shape[-1], block_k) == A_scale.shape[-1] + return A, A_scale + + +@triton.jit +def write_zeros_to_output( + c_ptr, + stride_cm, + stride_cn, + pid_n, + N, + offs_token, + token_mask, + BLOCK_SIZE_M, + BLOCK_SIZE_N, + compute_type, +): + accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=compute_type) + offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N) + c_ptrs = c_ptr + stride_cm * offs_token[:, None] + stride_cn * offs_cn[None, :] + c_mask = token_mask[:, None] & (offs_cn[None, :] < N) + tl.store(c_ptrs, accumulator, mask=c_mask) + + +@triton.jit +def fused_moe_kernel_gptq_awq( + # Pointers to matrices + a_ptr, + b_ptr, + c_ptr, + b_scale_ptr, + b_zp_ptr, + topk_weights_ptr, + sorted_token_ids_ptr, + expert_ids_ptr, + num_tokens_post_padded_ptr, + # Matrix dimensions + N: tl.constexpr, + K: tl.constexpr, + EM, + num_valid_tokens, + # The stride variables represent how much to increase the ptr by when + # moving by 1 element in a particular dimension. E.g. `stride_am` is + # how much to increase `a_ptr` by to get the element one row down + # (A has M rows). + stride_am, + stride_ak, + stride_be, + stride_bk, + stride_bn, + stride_cm, + stride_cn, + stride_bse, + stride_bsk, + stride_bsn, + stride_bze, + stride_bzk, + stride_bzn, + block_k_diviable: tl.constexpr, + group_size: tl.constexpr, + # Meta-parameters + BLOCK_SIZE_M: tl.constexpr, + BLOCK_SIZE_N: tl.constexpr, + BLOCK_SIZE_K: tl.constexpr, + GROUP_SIZE_M: tl.constexpr, + MUL_ROUTED_WEIGHT: tl.constexpr, + top_k: tl.constexpr, + compute_type: tl.constexpr, + has_zp: tl.constexpr, + use_int4_w4a16: tl.constexpr, + use_int8_w8a16: tl.constexpr, +): + """ + Implements the fused computation for a Mixture of Experts (MOE) using + token and expert matrices. + + Key Parameters: + - A: The input tensor representing tokens with shape (*, K), where '*' can + be any shape representing batches and K is the feature dimension of + each token. + - B: The stacked MOE weight tensor with shape (E, N, K), where E is + the number of experts, K is the input feature dimension, and N is + the output feature dimension. + - C: The output cache tensor with shape (M, topk, N), where M is the + total number of tokens post padding, topk is the number of times + each token is repeated, and N is the output feature dimension. + - sorted_token_ids: A tensor containing the sorted indices of tokens, + repeated topk times and arranged by the expert index they are + assigned to. + - expert_ids: A tensor containing the indices of the expert for each + block. It determines which expert matrix from B should be used for + each block in A. + This kernel performs the multiplication of a token by its corresponding + expert matrix as determined by `expert_ids`. The sorting of + `sorted_token_ids` by expert index and padding ensures divisibility by + BLOCK_SIZE_M, which is necessary to maintain consistency in block matrix + multiplication across different blocks processed by the same expert. + """ + # ----------------------------------------------------------- + # Map program ids `pid` to the block of C it should compute. + # This is done in a grouped ordering to promote L2 data reuse. + pid = tl.program_id(axis=0) + num_pid_m = tl.cdiv(EM, BLOCK_SIZE_M) + num_pid_n = tl.cdiv(N, BLOCK_SIZE_N) + num_pid_in_group = GROUP_SIZE_M * num_pid_n + group_id = pid // num_pid_in_group + first_pid_m = group_id * GROUP_SIZE_M + group_size_m = min(num_pid_m - first_pid_m, GROUP_SIZE_M) + pid_m = first_pid_m + ((pid % num_pid_in_group) % group_size_m) + pid_n = (pid % num_pid_in_group) // group_size_m + + # ---------------------------------------------------------- + # Create pointers for the first blocks of A and B. + # We will advance this pointer as we move in the K direction + # and accumulate + # `a_ptrs` is a block of [BLOCK_SIZE_M, BLOCK_SIZE_K] pointers + # `b_ptrs` is a block of [BLOCK_SIZE_K, BLOCK_SIZE_N] pointers + num_tokens_post_padded = tl.load(num_tokens_post_padded_ptr) + if pid_m * BLOCK_SIZE_M >= num_tokens_post_padded: + return + offs_token_id = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M).to(tl.int64) + offs_token = tl.load(sorted_token_ids_ptr + offs_token_id) + token_mask = offs_token < num_valid_tokens + + off_experts = tl.load(expert_ids_ptr + pid_m).to(tl.int64) + if off_experts == -1: + # ----------------------------------------------------------- + # Write back zeros to the output when the expert is not + # in the current expert parallel rank. + write_zeros_to_output( + c_ptr, + stride_cm, + stride_cn, + pid_n, + N, + offs_token, + token_mask, + BLOCK_SIZE_M, + BLOCK_SIZE_N, + compute_type, + ) + return + + offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N).to(tl.int64)) % N + offs_k = tl.arange(0, BLOCK_SIZE_K) + a_ptrs = a_ptr + ( + offs_token[:, None] // top_k * stride_am + offs_k[None, :] * stride_ak + ) + + if use_int4_w4a16: + b_ptrs = ( + b_ptr + + off_experts * stride_be + + (offs_k[:, None] // 2) * stride_bk + + offs_bn[None, :] * stride_bn + ) + b_shifter = (offs_k[:, None] % 2) * 4 + elif use_int8_w8a16: + b_ptrs = ( + b_ptr + + off_experts * stride_be + + offs_k[:, None] * stride_bk + + offs_bn[None, :] * stride_bn + ) + + if not has_zp and use_int4_w4a16: + b_zp_num = 8 + if not has_zp and use_int8_w8a16: + b_zp_num = 128 + elif has_zp and use_int4_w4a16: + b_zp_shifter = (offs_bn[None, :] % 2) * 4 + + # ----------------------------------------------------------- + # Iterate to compute a block of the C matrix. + # We accumulate into a `[BLOCK_SIZE_M, BLOCK_SIZE_N]` block + # of fp32 values for higher accuracy. + # `accumulator` will be converted back to fp16 after the loop. + accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32) + for k in range(0, tl.cdiv(K, BLOCK_SIZE_K)): + # Load the next block of A and B, generate a mask by checking the + # K dimension. + + if not block_k_diviable: + k_mask = offs_k[:, None] < K - k * BLOCK_SIZE_K + k_other = 0.0 + else: + k_mask = None + k_other = None + + a = tl.load( + a_ptrs, + mask=token_mask[:, None] & (offs_k[None, :] < K - k * BLOCK_SIZE_K), + other=0.0, + ) + b = tl.load(b_ptrs) + if use_int4_w4a16: + b = (b >> b_shifter) & 0xF + + b_scale_ptrs = ( + b_scale_ptr + + off_experts * stride_bse + + offs_bn[None, :] * stride_bsn + + ((offs_k[:, None] + BLOCK_SIZE_K * k) // group_size) * stride_bsk + ) + b_scale = tl.load(b_scale_ptrs, mask=k_mask, other=k_other) + b_scale = b_scale.to(tl.float32) + + if has_zp and use_int4_w4a16: + offs_k_true = (offs_k[:, None] + BLOCK_SIZE_K * k) // group_size + b_zp_ptrs = ( + b_zp_ptr + + off_experts * stride_bze + + (offs_bn[None, :] // 2) * stride_bzn + + offs_k_true * stride_bzk + ) + b_zp = tl.load(b_zp_ptrs, mask=k_mask, other=k_other) + b_zp = (b_zp >> b_zp_shifter) & 0xF + b_zp = b_zp.to(tl.float32) + elif has_zp and use_int8_w8a16: + offs_k_true = (offs_k[:, None] + BLOCK_SIZE_K * k) // group_size + b_zp_ptrs = ( + b_zp_ptr + + off_experts * stride_bze + + offs_bn[None, :] * stride_bzn + + offs_k_true * stride_bzk + ) + b_zp = tl.load(b_zp_ptrs, mask=k_mask, other=k_other) + b_zp = b_zp.to(tl.float32) + + # We accumulate along the K dimension. + if has_zp: + b = ((b.to(tl.float32) - b_zp) * b_scale).to(compute_type) + else: + b = ((b.to(tl.float32) - b_zp_num) * b_scale).to(compute_type) + accumulator = tl.dot(a, b, acc=accumulator) + + # Advance the ptrs to the next K block. + a_ptrs += BLOCK_SIZE_K * stride_ak + if use_int4_w4a16: + b_ptrs += (BLOCK_SIZE_K // 2) * stride_bk + else: + b_ptrs += BLOCK_SIZE_K * stride_bk + + if MUL_ROUTED_WEIGHT: + moe_weight = tl.load(topk_weights_ptr + offs_token, mask=token_mask, other=0) + accumulator = accumulator * moe_weight[:, None] + + accumulator = accumulator.to(compute_type) + # ----------------------------------------------------------- + # Write back the block of the output + offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N) + c_ptrs = c_ptr + stride_cm * offs_token[:, None] + stride_cn * offs_cn[None, :] + c_mask = token_mask[:, None] & (offs_cn[None, :] < N) + tl.store(c_ptrs, accumulator, mask=c_mask) + + +@triton.jit +def fused_moe_kernel( + # Pointers to matrices + a_ptr, + b_ptr, + c_ptr, + a_scale_ptr, + b_scale_ptr, + topk_weights_ptr, + sorted_token_ids_ptr, + expert_ids_ptr, + num_tokens_post_padded_ptr, + # Matrix dimensions + N, + K, + EM, + num_valid_tokens, + # The stride variables represent how much to increase the ptr by when + # moving by 1 element in a particular dimension. E.g. `stride_am` is + # how much to increase `a_ptr` by to get the element one row down + # (A has M rows). + stride_am, + stride_ak, + stride_be, + stride_bk, + stride_bn, + stride_cm, + stride_cn, + stride_asm, + stride_ask, + stride_bse, + stride_bsk, + stride_bsn, + # Block size for block-wise quantization + group_n: tl.constexpr, + group_k: tl.constexpr, + # Meta-parameters + BLOCK_SIZE_M: tl.constexpr, + BLOCK_SIZE_N: tl.constexpr, + BLOCK_SIZE_K: tl.constexpr, + GROUP_SIZE_M: tl.constexpr, + MUL_ROUTED_WEIGHT: tl.constexpr, + top_k: tl.constexpr, + compute_type: tl.constexpr, + use_fp8_w8a8: tl.constexpr, + use_int8_w8a16: tl.constexpr, +): + """ + Implements the fused computation for a Mixture of Experts (MOE) using + token and expert matrices. + + Key Parameters: + - A: The input tensor representing tokens with shape (*, K), where '*' can + be any shape representing batches and K is the feature dimension of + each token. + - B: The stacked MOE weight tensor with shape (E, N, K), where E is + the number of experts, K is the input feature dimension, and N is + the output feature dimension. + - C: The output cache tensor with shape (M, topk, N), where M is the + total number of tokens post padding, topk is the number of times + each token is repeated, and N is the output feature dimension. + - sorted_token_ids: A tensor containing the sorted indices of tokens, + repeated topk times and arranged by the expert index they are + assigned to. + - expert_ids: A tensor containing the indices of the expert for each + block. It determines which expert matrix from B should be used for + each block in A. + This kernel performs the multiplication of a token by its corresponding + expert matrix as determined by `expert_ids`. The sorting of + `sorted_token_ids` by expert index and padding ensures divisibility by + BLOCK_SIZE_M, which is necessary to maintain consistency in block matrix + multiplication across different blocks processed by the same expert. + """ + # ----------------------------------------------------------- + # Map program ids `pid` to the block of C it should compute. + # This is done in a grouped ordering to promote L2 data reuse. + pid = tl.program_id(axis=0) + num_pid_m = tl.cdiv(EM, BLOCK_SIZE_M) + num_pid_n = tl.cdiv(N, BLOCK_SIZE_N) + num_pid_in_group = GROUP_SIZE_M * num_pid_n + group_id = pid // num_pid_in_group + first_pid_m = group_id * GROUP_SIZE_M + group_size_m = min(num_pid_m - first_pid_m, GROUP_SIZE_M) + pid_m = first_pid_m + ((pid % num_pid_in_group) % group_size_m) + pid_n = (pid % num_pid_in_group) // group_size_m + + # ---------------------------------------------------------- + # Create pointers for the first blocks of A and B. + # We will advance this pointer as we move in the K direction + # and accumulate + # `a_ptrs` is a block of [BLOCK_SIZE_M, BLOCK_SIZE_K] pointers + # `b_ptrs` is a block of [BLOCK_SIZE_K, BLOCK_SIZE_N] pointers + num_tokens_post_padded = tl.load(num_tokens_post_padded_ptr) + if pid_m * BLOCK_SIZE_M >= num_tokens_post_padded: + return + offs_token_id = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M).to(tl.int64) + offs_token = tl.load(sorted_token_ids_ptr + offs_token_id) + token_mask = offs_token < num_valid_tokens + + off_experts = tl.load(expert_ids_ptr + pid_m).to(tl.int64) + if off_experts == -1: + # ----------------------------------------------------------- + # Write back zeros to the output when the expert is not + # in the current expert parallel rank. + write_zeros_to_output( + c_ptr, + stride_cm, + stride_cn, + pid_n, + N, + offs_token, + token_mask, + BLOCK_SIZE_M, + BLOCK_SIZE_N, + compute_type, + ) + return + + offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N).to(tl.int64)) % N + offs_k = tl.arange(0, BLOCK_SIZE_K) + a_ptrs = a_ptr + ( + offs_token[:, None] // top_k * stride_am + offs_k[None, :] * stride_ak + ) + + b_ptrs = ( + b_ptr + + off_experts * stride_be + + (offs_k[:, None] * stride_bk + offs_bn[None, :] * stride_bn) + ) + if use_int8_w8a16: + b_scale_ptrs = ( + b_scale_ptr + off_experts * stride_bse + offs_bn[None, :] * stride_bsn + ) + b_scale = tl.load(b_scale_ptrs) + + if use_fp8_w8a8: + if group_k > 0 and group_n > 0: + a_scale_ptrs = a_scale_ptr + (offs_token // top_k) * stride_asm + offs_bsn = offs_bn // group_n + b_scale_ptrs = ( + b_scale_ptr + off_experts * stride_bse + offs_bsn * stride_bsn + ) + else: + a_scale = tl.load(a_scale_ptr) + b_scale = tl.load(b_scale_ptr + off_experts) + + # ----------------------------------------------------------- + # Iterate to compute a block of the C matrix. + # We accumulate into a `[BLOCK_SIZE_M, BLOCK_SIZE_N]` block + # of fp32 values for higher accuracy. + # `accumulator` will be converted back to fp16 after the loop. + accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32) + for k in range(0, tl.cdiv(K, BLOCK_SIZE_K)): + # Load the next block of A and B, generate a mask by checking the + # K dimension. + a = tl.load( + a_ptrs, + mask=token_mask[:, None] & (offs_k[None, :] < K - k * BLOCK_SIZE_K), + other=0.0, + ) + b = tl.load(b_ptrs, mask=offs_k[:, None] < K - k * BLOCK_SIZE_K, other=0.0) + # We accumulate along the K dimension. + if use_int8_w8a16: + accumulator = tl.dot(a, b.to(compute_type), acc=accumulator) + elif use_fp8_w8a8: + if group_k > 0 and group_n > 0: + k_start = k * BLOCK_SIZE_K + offs_ks = k_start // group_k + a_scale = tl.load( + a_scale_ptrs + offs_ks * stride_ask, mask=token_mask, other=0.0 + ) + b_scale = tl.load(b_scale_ptrs + offs_ks * stride_bsk) + + accumulator += tl.dot(a, b) * a_scale[:, None] * b_scale[None, :] + else: + accumulator = tl.dot(a, b, acc=accumulator) + else: + accumulator += tl.dot(a, b) + # Advance the ptrs to the next K block. + a_ptrs += BLOCK_SIZE_K * stride_ak + b_ptrs += BLOCK_SIZE_K * stride_bk + + if MUL_ROUTED_WEIGHT: + moe_weight = tl.load(topk_weights_ptr + offs_token, mask=token_mask, other=0) + accumulator = accumulator * moe_weight[:, None] + if use_int8_w8a16: + accumulator = (accumulator * b_scale).to(compute_type) + elif use_fp8_w8a8: + if group_k > 0 and group_n > 0: + accumulator = accumulator.to(compute_type) + else: + accumulator = (accumulator * a_scale * b_scale).to(compute_type) + else: + accumulator = accumulator.to(compute_type) + # ----------------------------------------------------------- + # Write back the block of the output + offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N) + c_ptrs = c_ptr + stride_cm * offs_token[:, None] + stride_cn * offs_cn[None, :] + c_mask = token_mask[:, None] & (offs_cn[None, :] < N) + tl.store(c_ptrs, accumulator, mask=c_mask) + + +def invoke_fused_moe_kernel( + A: torch.Tensor, + B: torch.Tensor, + C: torch.Tensor, + A_scale: Optional[torch.Tensor], + B_scale: Optional[torch.Tensor], + B_zp: Optional[torch.Tensor], + topk_weights: Optional[torch.Tensor], + sorted_token_ids: torch.Tensor, + expert_ids: torch.Tensor, + num_tokens_post_padded: torch.Tensor, + mul_routed_weight: bool, + top_k: int, + config: Dict[str, Any], + compute_type: tl.dtype, + use_fp8_w8a8: bool, + use_int8_w8a16: bool, + use_int4_w4a16: bool, + block_shape: Optional[List[int]] = None, +) -> None: + assert topk_weights is not None or not mul_routed_weight + assert topk_weights is None or topk_weights.stride(1) == 1 + assert sorted_token_ids.stride(0) == 1 + + if use_fp8_w8a8: + assert B_scale is not None + assert ( + block_shape is None + or triton.cdiv(B.shape[-2], block_shape[0]) == B_scale.shape[-2] + ) + assert ( + block_shape is None + or triton.cdiv(B.shape[-1], block_shape[1]) == B_scale.shape[-1] + ) + + elif use_int8_w8a16 or use_int4_w4a16: + assert B_scale is not None + assert block_shape is None or block_shape[0] == 0 + else: + assert A_scale is None + assert B_scale is None + + M = A.shape[0] + num_tokens = M * top_k + + EM = sorted_token_ids.shape[0] + if A.shape[0] < config["BLOCK_SIZE_M"]: + # optimize for small batch_size. + # We assume that top_ids of each token is unique, so + # so num_valid_experts <= batch_size <= BLOCK_SIZE_M, + # and we can skip some invalid blocks. + EM = min(sorted_token_ids.shape[0], A.shape[0] * top_k * config["BLOCK_SIZE_M"]) + grid = lambda META: ( + triton.cdiv(EM, META["BLOCK_SIZE_M"]) + * triton.cdiv(B.shape[1], META["BLOCK_SIZE_N"]), + ) + + if ( + (use_int8_w8a16 or use_int4_w4a16) + and block_shape is not None + and block_shape[1] > 0 + ): + assert B_scale is not None and B_scale.ndim == 3 + assert B_zp is None or B_zp.ndim == 3 + + use_moe_wna16_cuda = should_moe_wna16_use_cuda( + num_valid_tokens=num_tokens, + group_size=block_shape[1], + num_experts=B.shape[0], + bit=4 if use_int4_w4a16 else 8, + ) + config = config.copy() + config.update( + get_moe_wna16_block_config( + config=config, + use_moe_wna16_cuda=use_moe_wna16_cuda, + num_valid_tokens=num_tokens, + size_k=A.shape[1], + size_n=B.shape[1], + num_experts=B.shape[1], + group_size=block_shape[1], + real_top_k=top_k, + block_size_m=config["BLOCK_SIZE_M"], + ) + ) + + if use_moe_wna16_cuda: + bit = 4 if use_int4_w4a16 else 8 + ops.moe_wna16_gemm( + A, + C, + B, + B_scale, + B_zp, + topk_weights if mul_routed_weight else None, + sorted_token_ids, + expert_ids, + num_tokens_post_padded, + top_k, + config["BLOCK_SIZE_M"], + config["BLOCK_SIZE_N"], + config["BLOCK_SIZE_K"], + bit, + ) + return + + fused_moe_kernel_gptq_awq[grid]( + A, + B, + C, + B_scale, + B_zp, + topk_weights, + sorted_token_ids, + expert_ids, + num_tokens_post_padded, + B.shape[1], + A.shape[1], + EM, + num_tokens, + A.stride(0), + A.stride(1), + B.stride(0), + B.stride(2), + B.stride(1), + C.stride(1), + C.stride(2), + B_scale.stride(0), + B_scale.stride(2), + B_scale.stride(1), + B_zp.stride(0) if B_zp is not None else 0, + B_zp.stride(2) if B_zp is not None else 0, + B_zp.stride(1) if B_zp is not None else 0, + block_k_diviable=A.shape[1] % config["BLOCK_SIZE_K"] == 0, + group_size=block_shape[1], + MUL_ROUTED_WEIGHT=mul_routed_weight, + top_k=top_k, + compute_type=compute_type, + has_zp=B_zp is not None, + use_int4_w4a16=use_int4_w4a16, + use_int8_w8a16=use_int8_w8a16, + **config, + ) + else: + config = config.copy() + BLOCK_SIZE_K = config.pop("BLOCK_SIZE_K") + if block_shape is not None: + BLOCK_SIZE_K = min(BLOCK_SIZE_K, min(block_shape[0], block_shape[1])) + fused_moe_kernel[grid]( + A, + B, + C, + A_scale, + B_scale, + topk_weights, + sorted_token_ids, + expert_ids, + num_tokens_post_padded, + B.shape[1], + B.shape[2], + EM, + num_tokens, + A.stride(0), + A.stride(1), + B.stride(0), + B.stride(2), + B.stride(1), + C.stride(1), + C.stride(2), + A_scale.stride(0) if A_scale is not None and A_scale.ndim == 2 else 0, + A_scale.stride(1) if A_scale is not None and A_scale.ndim == 2 else 0, + B_scale.stride(0) if B_scale is not None and B_scale.ndim >= 2 else 0, + B_scale.stride(2) if B_scale is not None and B_scale.ndim == 3 else 0, + B_scale.stride(1) if B_scale is not None and B_scale.ndim >= 2 else 0, + 0 if block_shape is None else block_shape[0], + 0 if block_shape is None else block_shape[1], + MUL_ROUTED_WEIGHT=mul_routed_weight, + top_k=top_k, + compute_type=compute_type, + use_fp8_w8a8=use_fp8_w8a8, + use_int8_w8a16=use_int8_w8a16, + BLOCK_SIZE_K=BLOCK_SIZE_K, + **config, + ) + + +# Adapted from: https://github.com/sgl-project/sglang/pull/2628 +def get_config_file_name( + E: int, N: int, dtype: Optional[str], block_shape: Optional[List[int]] = None +) -> str: + device_name = current_platform.get_device_name().replace(" ", "_") + dtype_selector = "" if not dtype else f",dtype={dtype}" + block_shape_selector = ( + "" if not block_shape or not all(block_shape) else f",block_shape={block_shape}" + ).replace(" ", "") + return f"E={E},N={N},device_name={device_name}{dtype_selector}{block_shape_selector}.json" # noqa: E501 + + +# Adapted from: https://github.com/sgl-project/sglang/pull/2628 +@functools.lru_cache +def get_moe_configs( + E: int, + N: int, + dtype: Optional[str], + block_n: Optional[int] = None, + block_k: Optional[int] = None, +) -> Optional[Dict[int, Any]]: + """ + Return optimized configurations for the fused MoE kernel. + + The return value will be a dictionary that maps an irregular grid of + batch sizes to configurations of the fused_moe kernel. To evaluate the + kernel on a given batch size bs, the closest batch size in the grid should + be picked and the associated configuration chosen to invoke the kernel. + """ + + # First look up if an optimized configuration is available in the configs + # directory + block_shape = [block_n, block_k] if block_n and block_k else None + json_file_name = get_config_file_name(E, N, dtype, block_shape) + + config_file_path = os.path.join( + os.path.dirname(os.path.realpath(__file__)), "configs", json_file_name + ) + if os.path.exists(config_file_path): + with open(config_file_path) as f: + logger.info("Using configuration from %s for MoE layer.", config_file_path) + # If a configuration has been found, return it + return {int(key): val for key, val in json.load(f).items()} + + # If no optimized configuration is available, we will use the default + # configuration + logger.warning( + ( + "Using default MoE config. Performance might be sub-optimal! " + "Config file not found at %s" + ), + config_file_path, + ) + return None + + +def get_moe_wna16_block_config( + config: Dict[str, int], + use_moe_wna16_cuda: bool, + num_valid_tokens: int, + size_k: int, + size_n: int, + num_experts: int, + group_size: int, + real_top_k: int, + block_size_m: int, +): + if "BLOCK_SIZE_N" in config and "BLOCK_SIZE_K" in config: + # optimal block config is set + return {} + if not use_moe_wna16_cuda: + # triton moe wna16 kernel + if num_valid_tokens // real_top_k == 1: + # if bs=1, use a smaller BLOCK_SIZE_N + return {"BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64} + else: + return {"BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32} + else: + # cuda moe wna16 kernel + # set default block_size 128, and increase them when num_blocks + # is too large. + block_size_n = 128 + block_size_k = 128 + if block_size_k <= group_size: + block_size_k = group_size + + num_n_blocks = size_k // block_size_k + num_k_blocks = size_n // block_size_k + num_m_blocks = ( + num_valid_tokens + block_size_m - 1 + ) / block_size_m + num_experts + if num_valid_tokens // real_top_k <= block_size_m: + num_m_blocks = min(num_m_blocks, num_valid_tokens) + num_blocks = num_m_blocks * num_n_blocks * num_k_blocks + + if size_k % 256 == 0 and num_blocks >= 256 and block_size_k < 256: + block_size_k = 256 + num_blocks = num_blocks // (256 // block_size_k) + + if ( + num_m_blocks <= 16 + and size_k % (block_size_k * 2) == 0 + and size_k % (block_size_k * 2) == 0 + and block_size_k <= 512 + and num_blocks >= 512 + ): + block_size_k = block_size_k * 2 + num_blocks = num_blocks // 2 + + if num_blocks > 1024: + block_size_n = 256 + num_n_blocks = num_n_blocks // 2 + num_blocks = num_blocks // 2 + + if size_n <= 1024 and num_blocks >= 1024: + # The kernel performance got much better with BLOCK_SIZE_N=1024 + # when num_blocks is large, event when N is small. + # Not sure why, maybe it force the CUDA SM process only one block + # at the same time. + block_size_n = 1024 + + return {"BLOCK_SIZE_N": block_size_n, "BLOCK_SIZE_K": block_size_k} + + +def should_moe_wna16_use_cuda( + num_valid_tokens: int, group_size: int, num_experts: int, bit: int +): + return ( + bit == 4 and group_size in [32, 64, 128] and num_valid_tokens / num_experts <= 6 + ) + + +def get_default_config( + M: int, + E: int, + N: int, + K: int, + topk: int, + dtype: Optional[str], + is_marlin: bool, + block_shape: Optional[List[int]] = None, +) -> Dict[str, int]: + if dtype == "fp8_w8a8" and block_shape is not None: + # Block-wise quant: BLOCK_SIZE_N must be divisible by block_shape[0] + # BLOCK_SIZE_K must be divisible by block_shape[1] + config = { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": block_shape[0], + "BLOCK_SIZE_K": block_shape[1], + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3, + } + elif dtype in ["int4_w4a16", "int8_w8a16"] and block_shape is not None: + # moe wna16 kernels + # only set BLOCK_SIZE_M + # BLOCK_SIZE_N and BLOCK_SIZE_K would be set later + bit = 4 if dtype == "int4_w4a16" else 8 + use_moe_wna16_cuda = should_moe_wna16_use_cuda(M * topk, block_shape[1], E, bit) + if use_moe_wna16_cuda: + config = {"BLOCK_SIZE_M": min(16, M)} + elif M <= 20: + config = {"BLOCK_SIZE_M": 16, "GROUP_SIZE_M": 1} + elif M <= 40: + config = {"BLOCK_SIZE_M": 32, "GROUP_SIZE_M": 1} + else: + config = {"BLOCK_SIZE_M": 64, "GROUP_SIZE_M": 1} + else: + config = { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 8, + } + # A heuristic: fused marlin works faster with this config for small M + if M <= E or (is_marlin and M <= 32): + config = { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + } + return config + + +def try_get_optimal_moe_config( + w1_shape: Tuple[int, ...], + w2_shape: Tuple[int, ...], + top_k: int, + dtype: Optional[str], + M: int, + is_marlin: bool = False, + block_shape: Optional[List[int]] = None, +): + # from vllm.model_executor.layers.fused_moe import get_config + # TODO: removed when syncing to vLLM, do we need this? + # override_config = get_config() + override_config = None + if override_config: + config = override_config + else: + # First try to load optimal config from the file + E, _, N = w2_shape + if dtype == "int4_w4a16": + N = N * 2 + block_n = block_shape[0] if block_shape else 0 + block_k = block_shape[1] if block_shape else 0 + configs = get_moe_configs(E, N, dtype, block_n, block_k) + + if configs: + # If an optimal configuration map has been found, look up the + # optimal config + config = configs[min(configs.keys(), key=lambda x: abs(x - M))] + else: + # Else use the default config + config = get_default_config( + M, E, N, w1_shape[2], top_k, dtype, is_marlin, block_shape + ) + return config + + +def vllm_topk_softmax( + topk_weights: torch.Tensor, + topk_indices: torch.Tensor, + token_expert_indices: torch.Tensor, + gating_output: torch.Tensor, + renormalize: bool, +) -> tuple[torch.Tensor, ...]: + ops.topk_softmax( + topk_weights, + topk_indices, + token_expert_indices, + gating_output, + ) + if renormalize: + topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True) + + return topk_weights, topk_indices + + +def dispatch_topk_func() -> Callable[..., tuple[torch.Tensor, ...]]: + return vllm_topk_softmax + + +def fused_topk( + hidden_states: torch.Tensor, + gating_output: torch.Tensor, + topk: int, + renormalize: bool, +) -> Tuple[torch.Tensor, torch.Tensor]: + assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch" + + M, _ = hidden_states.shape + + topk_weights = torch.empty( + M, topk, dtype=torch.float32, device=hidden_states.device + ) + topk_ids = torch.empty(M, topk, dtype=torch.int32, device=hidden_states.device) + token_expert_indicies = torch.empty( + M, topk, dtype=torch.int32, device=hidden_states.device + ) + + gating_output_float = gating_output.float() # TODO(woosuk): Optimize this. + + topk_func = dispatch_topk_func() + topk_weights, topk_ids = topk_func( + topk_weights, topk_ids, token_expert_indicies, gating_output_float, renormalize + ) + + del token_expert_indicies # Not used. Will be used in the future. + return topk_weights, topk_ids + + +# This is used by the Deepseek-V2 and Deepseek-V3 model +@torch.compile(dynamic=True, backend=current_platform.simple_compile_backend) +def grouped_topk( + hidden_states: torch.Tensor, + gating_output: torch.Tensor, + topk: int, + renormalize: bool, + num_expert_group: int = 0, + topk_group: int = 0, + scoring_func: str = "softmax", + e_score_correction_bias: Optional[torch.Tensor] = None, +) -> Tuple[torch.Tensor, torch.Tensor]: + + assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch" + + if scoring_func == "softmax": + scores = torch.softmax(gating_output, dim=-1) + elif scoring_func == "sigmoid": + scores = gating_output.sigmoid() + else: + raise ValueError(f"Unsupported scoring function: {scoring_func}") + + num_token = scores.shape[0] + if e_score_correction_bias is not None: + # Store original scores before applying correction bias. We use biased + # scores for expert selection but original scores for routing weights + original_scores = scores + scores = scores + e_score_correction_bias.unsqueeze(0) + group_scores = ( + scores.view(num_token, num_expert_group, -1).topk(2, dim=-1)[0].sum(dim=-1) + ) + else: + group_scores = ( + scores.view(num_token, num_expert_group, -1).max(dim=-1).values + ) # [n, n_group] + group_idx = torch.topk(group_scores, k=topk_group, dim=-1, sorted=False)[ + 1 + ] # [n, top_k_group] + group_mask = torch.zeros_like(group_scores) # [n, n_group] + group_mask.scatter_(1, group_idx, 1) # [n, n_group] + score_mask = ( + group_mask.unsqueeze(-1) + .expand(num_token, num_expert_group, scores.shape[-1] // num_expert_group) + .reshape(num_token, -1) + ) # [n, e] + tmp_scores = scores.masked_fill(~score_mask.bool(), float("-inf")) # [n, e] + + if e_score_correction_bias is not None: + topk_ids = torch.topk(tmp_scores, k=topk, dim=-1, sorted=False)[1] + # Use original unbiased scores for the routing weights + topk_weights = original_scores.gather(1, topk_ids) + else: + topk_weights, topk_ids = torch.topk(tmp_scores, k=topk, dim=-1, sorted=False) + + if renormalize: + topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True) + + return topk_weights.to(torch.float32), topk_ids.to(torch.int32) + + +def get_config_dtype_str( + dtype: torch.dtype, + use_int4_w4a16: Optional[bool] = False, + use_int8_w8a16: Optional[bool] = False, + use_fp8_w8a8: Optional[bool] = False, +) -> Optional[str]: + if use_fp8_w8a8: + return "fp8_w8a8" + elif use_int8_w8a16: + return "int8_w8a16" + elif use_int4_w4a16: + return "int4_w4a16" + elif dtype == torch.float: + # avoiding cases where kernel fails when float32 MoE + # use fp16/bfloat16 configs + return "float32" + return None + + +def inplace_fused_experts( + hidden_states: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + activation: str = "silu", + apply_router_weight_on_input: bool = False, + use_fp8_w8a8: bool = False, + use_int8_w8a16: bool = False, + use_int4_w4a16: bool = False, + global_num_experts: int = -1, + expert_map: Optional[torch.Tensor] = None, + w1_scale: Optional[torch.Tensor] = None, + w2_scale: Optional[torch.Tensor] = None, + w1_zp: Optional[torch.Tensor] = None, + w2_zp: Optional[torch.Tensor] = None, + a1_scale: Optional[torch.Tensor] = None, + a2_scale: Optional[torch.Tensor] = None, + block_shape: Optional[List[int]] = None, +) -> None: + fused_experts_impl( + hidden_states, + w1, + w2, + topk_weights, + topk_ids, + True, + activation, + apply_router_weight_on_input, + use_fp8_w8a8, + use_int8_w8a16, + use_int4_w4a16, + global_num_experts, + expert_map, + w1_scale, + w2_scale, + w1_zp, + w2_zp, + a1_scale, + a2_scale, + block_shape, + ) + + +def inplace_fused_experts_fake( + hidden_states: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + activation: str = "silu", + apply_router_weight_on_input: bool = False, + use_fp8_w8a8: bool = False, + use_int8_w8a16: bool = False, + use_int4_w4a16: bool = False, + global_num_experts: int = -1, + expert_map: Optional[torch.Tensor] = None, + w1_scale: Optional[torch.Tensor] = None, + w2_scale: Optional[torch.Tensor] = None, + w1_zp: Optional[torch.Tensor] = None, + w2_zp: Optional[torch.Tensor] = None, + a1_scale: Optional[torch.Tensor] = None, + a2_scale: Optional[torch.Tensor] = None, + block_shape: Optional[List[int]] = None, +) -> None: + pass + + +direct_register_custom_op( + op_name="inplace_fused_experts", + op_func=inplace_fused_experts, + mutates_args=["hidden_states"], + fake_impl=inplace_fused_experts_fake, +) + + +def outplace_fused_experts( + hidden_states: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + activation: str = "silu", + apply_router_weight_on_input: bool = False, + use_fp8_w8a8: bool = False, + use_int8_w8a16: bool = False, + use_int4_w4a16: bool = False, + global_num_experts: int = -1, + expert_map: Optional[torch.Tensor] = None, + w1_scale: Optional[torch.Tensor] = None, + w2_scale: Optional[torch.Tensor] = None, + w1_zp: Optional[torch.Tensor] = None, + w2_zp: Optional[torch.Tensor] = None, + a1_scale: Optional[torch.Tensor] = None, + a2_scale: Optional[torch.Tensor] = None, + block_shape: Optional[List[int]] = None, +) -> torch.Tensor: + return fused_experts_impl( + hidden_states, + w1, + w2, + topk_weights, + topk_ids, + False, + activation, + apply_router_weight_on_input, + use_fp8_w8a8, + use_int8_w8a16, + use_int4_w4a16, + global_num_experts, + expert_map, + w1_scale, + w2_scale, + w1_zp, + w2_zp, + a1_scale, + a2_scale, + block_shape, + ) + + +def outplace_fused_experts_fake( + hidden_states: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + activation: str = "silu", + apply_router_weight_on_input: bool = False, + use_fp8_w8a8: bool = False, + use_int8_w8a16: bool = False, + use_int4_w4a16: bool = False, + global_num_experts: int = -1, + expert_map: Optional[torch.Tensor] = None, + w1_scale: Optional[torch.Tensor] = None, + w2_scale: Optional[torch.Tensor] = None, + w1_zp: Optional[torch.Tensor] = None, + w2_zp: Optional[torch.Tensor] = None, + a1_scale: Optional[torch.Tensor] = None, + a2_scale: Optional[torch.Tensor] = None, + block_shape: Optional[List[int]] = None, +) -> torch.Tensor: + return torch.empty_like(hidden_states) + + +direct_register_custom_op( + op_name="outplace_fused_experts", + op_func=outplace_fused_experts, + mutates_args=[], + fake_impl=outplace_fused_experts_fake, +) + + +def torch_vllm_inplace_fused_experts(**kwargs) -> torch.Tensor: + ops.inplace_fused_experts(**kwargs) + hidden_states = kwargs["hidden_states"] + return hidden_states + + +def torch_vllm_outplace_fused_experts(**kwargs) -> torch.Tensor: + return ops.outplace_fused_experts(**kwargs) + + +def dispatch_fused_experts_func(inplace: bool) -> Callable[..., torch.Tensor]: + if inplace: + return torch_vllm_inplace_fused_experts + return torch_vllm_outplace_fused_experts + + +def fused_experts( + hidden_states: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + inplace: bool = False, + activation: str = "silu", + apply_router_weight_on_input: bool = False, + use_fp8_w8a8: bool = False, + use_int8_w8a16: bool = False, + use_int4_w4a16: bool = False, + global_num_experts: int = -1, + expert_map: Optional[torch.Tensor] = None, + w1_scale: Optional[torch.Tensor] = None, + w2_scale: Optional[torch.Tensor] = None, + w1_zp: Optional[torch.Tensor] = None, + w2_zp: Optional[torch.Tensor] = None, + a1_scale: Optional[torch.Tensor] = None, + a2_scale: Optional[torch.Tensor] = None, + block_shape: Optional[List[int]] = None, +) -> torch.Tensor: + return dispatch_fused_experts_func(inplace)( + hidden_states=hidden_states, + w1=w1, + w2=w2, + topk_weights=topk_weights, + topk_ids=topk_ids, + activation=activation, + apply_router_weight_on_input=apply_router_weight_on_input, + use_fp8_w8a8=use_fp8_w8a8, + use_int8_w8a16=use_int8_w8a16, + use_int4_w4a16=use_int4_w4a16, + global_num_experts=global_num_experts, + expert_map=expert_map, + w1_scale=w1_scale, + w2_scale=w2_scale, + w1_zp=w1_zp, + w2_zp=w2_zp, + a1_scale=a1_scale, + a2_scale=a2_scale, + block_shape=block_shape, + ) + + +def fused_experts_impl( + hidden_states: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + inplace: bool = False, + activation: str = "silu", + apply_router_weight_on_input: bool = False, + use_fp8_w8a8: bool = False, + use_int8_w8a16: bool = False, + use_int4_w4a16: bool = False, + global_num_experts: int = -1, + expert_map: Optional[torch.Tensor] = None, + w1_scale: Optional[torch.Tensor] = None, + w2_scale: Optional[torch.Tensor] = None, + w1_zp: Optional[torch.Tensor] = None, + w2_zp: Optional[torch.Tensor] = None, + a1_scale: Optional[torch.Tensor] = None, + a2_scale: Optional[torch.Tensor] = None, + block_shape: Optional[List[int]] = None, +): + # Check constraints. + if use_int4_w4a16: + assert hidden_states.shape[1] // 2 == w1.shape[2], "Hidden size mismatch" + else: + assert hidden_states.shape[1] == w1.shape[2], "Hidden size mismatch" + + assert topk_weights.shape == topk_ids.shape, "topk shape mismatch" + assert hidden_states.is_contiguous(), "Hidden_states must be contiguous" + assert w1.stride(-1) == 1, "Stride of last dimension must be 1" + assert w2.stride(-1) == 1, "Stride of last dimension must be 1" + assert hidden_states.dtype in [torch.float32, torch.float16, torch.bfloat16] + + num_tokens, _ = hidden_states.shape + E, N, _ = w1.shape + K = w2.shape[1] + if global_num_experts == -1: + global_num_experts = E + top_k_num = topk_ids.shape[1] + # We execute the fused_moe kernel in chunks to circumvent this issue: + # https://github.com/vllm-project/vllm/issues/5938 + CHUNK_SIZE = VLLM_FUSED_MOE_CHUNK_SIZE + M = min(num_tokens, CHUNK_SIZE) + config_dtype = get_config_dtype_str( + use_fp8_w8a8=use_fp8_w8a8, + use_int8_w8a16=use_int8_w8a16, + use_int4_w4a16=use_int4_w4a16, + dtype=hidden_states.dtype, + ) + + get_config_func = functools.partial( + try_get_optimal_moe_config, + w1.shape, + w2.shape, + top_k_num, + config_dtype, + block_shape=block_shape, + ) + + config = get_config_func(M) + + # We can reuse the memory between these because by the time we need + # cache3, we're done with cache1 + cache13 = torch.empty( + M * top_k_num * max(N, K), + device=hidden_states.device, + dtype=hidden_states.dtype, + ) + intermediate_cache1 = cache13[: M * top_k_num * N].view(M, top_k_num, N) + intermediate_cache3 = cache13[: M * top_k_num * K].view(M, top_k_num, K) + + # This needs separate memory since it's used concurrently with cache1 + intermediate_cache2 = torch.empty( + (M * top_k_num, N // 2), device=hidden_states.device, dtype=hidden_states.dtype + ) + + if hidden_states.dtype == torch.bfloat16: + compute_type = tl.bfloat16 + elif hidden_states.dtype == torch.float16: + compute_type = tl.float16 + elif hidden_states.dtype == torch.float32: + compute_type = tl.float32 + else: + raise ValueError(f"Unsupported compute_type: {hidden_states.dtype}") + + if inplace: + out_hidden_states = hidden_states + else: + out_hidden_states = torch.empty_like(hidden_states) + + for chunk in range((num_tokens // CHUNK_SIZE) + 1): + begin_chunk_idx, end_chunk_idx = ( + chunk * CHUNK_SIZE, + min((chunk + 1) * CHUNK_SIZE, num_tokens), + ) + curr_hidden_states = hidden_states[begin_chunk_idx:end_chunk_idx] + tokens_in_chunk, _ = curr_hidden_states.shape + + if tokens_in_chunk == 0: + break + + if tokens_in_chunk < CHUNK_SIZE and chunk > 0: + # Adjust the intermediate cache size and config for the last + # chunk. Note that in most cases we only have one chunk + # so the cache size and config are already set correctly and + # do not need to be adjusted. + intermediate_cache1 = intermediate_cache1[:tokens_in_chunk] + intermediate_cache2 = intermediate_cache2[ + : tokens_in_chunk * topk_ids.shape[1] + ] + intermediate_cache3 = intermediate_cache3[:tokens_in_chunk] + config = get_config_func(tokens_in_chunk) + + curr_topk_ids = topk_ids[begin_chunk_idx:end_chunk_idx] + curr_topk_weights = topk_weights[begin_chunk_idx:end_chunk_idx] + + a1q_scale: Optional[torch.Tensor] = None + + if use_fp8_w8a8: + qcurr_hidden_states, a1q_scale = _fp8_quantize( + curr_hidden_states, a1_scale, block_shape + ) + else: + qcurr_hidden_states = curr_hidden_states + a1q_scale = a1_scale + + sorted_token_ids, expert_ids, num_tokens_post_padded = moe_align_block_size( + curr_topk_ids, config["BLOCK_SIZE_M"], global_num_experts, expert_map + ) + + invoke_fused_moe_kernel( + qcurr_hidden_states, + w1, + intermediate_cache1, + a1q_scale, + w1_scale, + w1_zp, + curr_topk_weights, + sorted_token_ids, + expert_ids, + num_tokens_post_padded, + apply_router_weight_on_input, + top_k_num, + config, + compute_type=compute_type, + use_fp8_w8a8=use_fp8_w8a8, + use_int8_w8a16=use_int8_w8a16, + use_int4_w4a16=use_int4_w4a16, + block_shape=block_shape, + ) + + if activation == "silu": + ops.silu_and_mul(intermediate_cache2, intermediate_cache1.view(-1, N)) + elif activation == "gelu": + ops.gelu_and_mul(intermediate_cache2, intermediate_cache1.view(-1, N)) + else: + raise ValueError(f"Unsupported FusedMoe activation: {activation}") + + a2q_scale: Optional[torch.Tensor] = None + + if use_fp8_w8a8: + qintermediate_cache2, a2q_scale = _fp8_quantize( + intermediate_cache2, a2_scale, block_shape + ) + else: + qintermediate_cache2 = intermediate_cache2 + a2q_scale = a2_scale + + invoke_fused_moe_kernel( + qintermediate_cache2, + w2, + intermediate_cache3, + a2q_scale, + w2_scale, + w2_zp, + curr_topk_weights, + sorted_token_ids, + expert_ids, + num_tokens_post_padded, + not apply_router_weight_on_input, + 1, + config, + compute_type=compute_type, + use_fp8_w8a8=use_fp8_w8a8, + use_int8_w8a16=use_int8_w8a16, + use_int4_w4a16=use_int4_w4a16, + block_shape=block_shape, + ) + + ops.moe_sum( + intermediate_cache3.view(*intermediate_cache3.shape), + out_hidden_states[begin_chunk_idx:end_chunk_idx], + ) + + return out_hidden_states + + +def fused_moe( + hidden_states: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + gating_output: torch.Tensor, + topk: int, + renormalize: bool, + inplace: bool = False, + activation: str = "silu", + apply_router_weight_on_input: bool = False, + use_grouped_topk: bool = False, + num_expert_group: Optional[int] = None, + topk_group: Optional[int] = None, + custom_routing_function: Optional[Callable] = None, + use_fp8_w8a8: bool = False, + use_int8_w8a16: bool = False, + use_int4_w4a16: bool = False, + global_num_experts: int = -1, + expert_map: Optional[torch.Tensor] = None, + w1_scale: Optional[torch.Tensor] = None, + w2_scale: Optional[torch.Tensor] = None, + w1_zp: Optional[torch.Tensor] = None, + w2_zp: Optional[torch.Tensor] = None, + a1_scale: Optional[torch.Tensor] = None, + a2_scale: Optional[torch.Tensor] = None, + block_shape: Optional[List[int]] = None, +) -> torch.Tensor: + """ + This function computes a Mixture of Experts (MoE) layer using two sets of + weights, w1 and w2, and top-k gating mechanism. + + Parameters: + - hidden_states (torch.Tensor): The input tensor to the MoE layer. + - w1 (torch.Tensor): The first set of expert weights. + - w2 (torch.Tensor): The second set of expert weights. + - gating_output (torch.Tensor): The output of the gating operation + (before softmax). + - topk (int): The number of top-k experts to select. + - renormalize (bool): If True, renormalize the top-k weights to sum to 1. + - inplace (bool): If True, perform the operation in-place. + Defaults to False. + - activation (str): The activation function to apply after the first + MoE layer. + - num_expert_group: Optional[int]: additional parameter for grouped_topk + - topk_group: Optional[int]: additional parameter for grouped_topk + - use_grouped_topk: If True, use grouped_topk instead of fused_topk + note: Deepseekv2 model uses grouped_topk + - use_fp8_w8a8 (bool): If True, use fp8 arithmetic to compute the inner + products for w1 and w2. Defaults to False. + - use_int8_w8a16 (bool): If True, use matmul of int8 weight and bf16/fp16 + activation to compute the inner products for w1 and w2. + Defaults to False. + - use_int4_w4a16 (bool): If True, use matmul of int4 weight and bf16/fp16 + activation to compute the inner products for w1 and w2. + Defaults to False. + - global_num_experts (int): The total number of experts in the global + expert space. + - expert_map (Optional[torch.Tensor]): A tensor mapping expert indices + from the global expert space to the local expert space of the expert + parallel shard. + - w1_scale (Optional[torch.Tensor]): Optional scale to be used for + w1. + - w2_scale (Optional[torch.Tensor]): Optional scale to be used for + w2. + - a1_scale (Optional[torch.Tensor]): Optional scale to be used for + a1. + - a2_scale (Optional[torch.Tensor]): Optional scale to be used for + a2. + - block_shape: (Optional[List[int]]): Optional block size for block-wise + quantization. + + Returns: + - torch.Tensor: The output tensor after applying the MoE layer. + """ + + if use_grouped_topk: + assert num_expert_group is not None and topk_group is not None + topk_weights, topk_ids = grouped_topk( + hidden_states, + gating_output, + topk, + renormalize, + num_expert_group, + topk_group, + ) + elif custom_routing_function is None: + topk_weights, topk_ids = fused_topk( + hidden_states, gating_output, topk, renormalize + ) + else: + topk_weights, topk_ids = custom_routing_function( + hidden_states, gating_output, topk, renormalize + ) + + return fused_experts( + hidden_states, + w1, + w2, + topk_weights, + topk_ids, + inplace=inplace, + activation=activation, + apply_router_weight_on_input=apply_router_weight_on_input, + use_fp8_w8a8=use_fp8_w8a8, + use_int8_w8a16=use_int8_w8a16, + use_int4_w4a16=use_int4_w4a16, + global_num_experts=global_num_experts, + expert_map=expert_map, + w1_scale=w1_scale, + w2_scale=w2_scale, + w1_zp=w1_zp, + w2_zp=w2_zp, + a1_scale=a1_scale, + a2_scale=a2_scale, + block_shape=block_shape, + ) diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/layers.py b/build/torch26-cxx11-cu126-aarch64-linux/moe/layers.py new file mode 100644 index 0000000000000000000000000000000000000000..1befea3f9e97a6a4aa1679821d30091c3d825bb8 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/layers.py @@ -0,0 +1,64 @@ +import torch +import torch.nn as nn + +from .fused_moe import fused_moe + + +def _llama4_topk( + hidden_states: torch.Tensor, + gating_output: torch.Tensor, + topk: int, + renormalize: bool, +): + assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch" + topk_weights, topk_ids = torch.topk(gating_output, topk, dim=1) + topk_weights = torch.sigmoid(topk_weights.float()).to(hidden_states.dtype) + return topk_weights, topk_ids + + +def _fix_llama4_experts(hidden_states: torch.Tensor, experts: nn.Module): + if experts.gate_up_proj.shape[-1] != hidden_states.shape[-1]: + experts.gate_up_proj = nn.Parameter( + experts.gate_up_proj.transpose(1, 2).contiguous() + ) + experts.down_proj = nn.Parameter(experts.down_proj.transpose(1, 2).contiguous()) + + +class Llama4TextMoe(nn.Module): + has_backward = False + + experts: nn.Module + router: nn.Linear + shared_expert: nn.Module + top_k: int + + def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: + hidden_states = hidden_states.reshape(-1, hidden_states.shape[-1]) + + _fix_llama4_experts(hidden_states, self.experts) + + router_logits = self.router(hidden_states) + + extra_kwargs = {} + use_fp8_w8a8 = False + if hasattr(self.experts, "gate_up_proj_scale"): + use_fp8_w8a8 = True + extra_kwargs["w1_scale"] = self.experts.gate_up_proj_scale + extra_kwargs["w2_scale"] = self.experts.down_proj_scale + + out = fused_moe( + hidden_states, + w1=self.experts.gate_up_proj, + w2=self.experts.down_proj, + gating_output=router_logits, + topk=self.top_k, + renormalize=False, + custom_routing_function=_llama4_topk, + apply_router_weight_on_input=True, + use_fp8_w8a8=use_fp8_w8a8, + **extra_kwargs + ) + + out += self.shared_expert(hidden_states) + + return out, router_logits.t() diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/moe_align_block_size.py b/build/torch26-cxx11-cu126-aarch64-linux/moe/moe_align_block_size.py new file mode 100644 index 0000000000000000000000000000000000000000..a698de5aaa123df5a2d663f252e1d5914e880d09 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/moe_align_block_size.py @@ -0,0 +1,252 @@ +# SPDX-License-Identifier: Apache-2.0 +import os +from typing import Optional, Tuple + +import torch +import triton +import triton.language as tl + +from ._ops import ops + + +def round_up(x: int, y: int) -> int: + return ((x + y - 1) // y) * y + + +VLLM_ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON = bool( + int(os.getenv("VLLM_ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON", "0")) +) + + +def ceil_div(a, b): + return (a + b - 1) // b + + +@triton.jit +def moe_align_block_size_stage1( + topk_ids_ptr, + tokens_cnts_ptr, + num_experts: tl.constexpr, + numel: tl.constexpr, + tokens_per_thread: tl.constexpr, +): + pid = tl.program_id(0) + + start_idx = pid * tokens_per_thread + + off_c = (pid + 1) * num_experts + + for i in range(tokens_per_thread): + if start_idx + i < numel: + idx = tl.load(topk_ids_ptr + start_idx + i) + token_cnt = tl.load(tokens_cnts_ptr + off_c + idx) + tl.store(tokens_cnts_ptr + off_c + idx, token_cnt + 1) + + +@triton.jit +def moe_align_block_size_stage2( + tokens_cnts_ptr, + num_experts: tl.constexpr, +): + pid = tl.program_id(0) + + last_cnt = 0 + for i in range(1, num_experts + 1): + token_cnt = tl.load(tokens_cnts_ptr + i * num_experts + pid) + last_cnt = last_cnt + token_cnt + tl.store(tokens_cnts_ptr + i * num_experts + pid, last_cnt) + + +@triton.jit +def moe_align_block_size_stage3( + total_tokens_post_pad_ptr, + tokens_cnts_ptr, + cumsum_ptr, + num_experts: tl.constexpr, + block_size: tl.constexpr, +): + last_cumsum = 0 + off_cnt = num_experts * num_experts + for i in range(1, num_experts + 1): + token_cnt = tl.load(tokens_cnts_ptr + off_cnt + i - 1) + last_cumsum = last_cumsum + tl.cdiv(token_cnt, block_size) * block_size + tl.store(cumsum_ptr + i, last_cumsum) + tl.store(total_tokens_post_pad_ptr, last_cumsum) + + +@triton.jit +def moe_align_block_size_stage4( + topk_ids_ptr, + sorted_token_ids_ptr, + expert_ids_ptr, + tokens_cnts_ptr, + cumsum_ptr, + num_experts: tl.constexpr, + block_size: tl.constexpr, + numel: tl.constexpr, + tokens_per_thread: tl.constexpr, +): + pid = tl.program_id(0) + start_idx = tl.load(cumsum_ptr + pid) + end_idx = tl.load(cumsum_ptr + pid + 1) + + for i in range(start_idx, end_idx, block_size): + tl.store(expert_ids_ptr + i // block_size, pid) + + start_idx = pid * tokens_per_thread + off_t = pid * num_experts + + for i in range(start_idx, tl.minimum(start_idx + tokens_per_thread, numel)): + expert_id = tl.load(topk_ids_ptr + i) + token_cnt = tl.load(tokens_cnts_ptr + off_t + expert_id) + rank_post_pad = token_cnt + tl.load(cumsum_ptr + expert_id) + tl.store(sorted_token_ids_ptr + rank_post_pad, i) + tl.store(tokens_cnts_ptr + off_t + expert_id, token_cnt + 1) + + +# Triton implementation based on: +# https://github.com/sgl-project/sglang/commit/ba5112ff691d791a9e38c6c71f59324a5fcb49d0 +def moe_align_block_size_triton( + topk_ids: torch.Tensor, + num_experts: int, + block_size: int, + sorted_token_ids: torch.Tensor, + expert_ids: torch.Tensor, + num_tokens_post_pad: torch.Tensor, +) -> None: + numel = topk_ids.numel() + grid = (num_experts,) + tokens_cnts = torch.zeros( + (num_experts + 1, num_experts), dtype=torch.int32, device=topk_ids.device + ) + cumsum = torch.zeros((num_experts + 1,), dtype=torch.int32, device=topk_ids.device) + tokens_per_thread = ceil_div(numel, num_experts) + + moe_align_block_size_stage1[grid]( + topk_ids, + tokens_cnts, + num_experts, + numel, + tokens_per_thread, + ) + moe_align_block_size_stage2[grid]( + tokens_cnts, + num_experts, + ) + moe_align_block_size_stage3[(1,)]( + num_tokens_post_pad, + tokens_cnts, + cumsum, + num_experts, + block_size, + ) + moe_align_block_size_stage4[grid]( + topk_ids, + sorted_token_ids, + expert_ids, + tokens_cnts, + cumsum, + num_experts, + block_size, + numel, + tokens_per_thread, + ) + + +def moe_align_block_size( + topk_ids: torch.Tensor, + block_size: int, + num_experts: int, + expert_map: Optional[torch.Tensor] = None, + pad_sorted_ids: bool = False, +) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + """ + Aligns the token distribution across experts to be compatible with block + size for matrix multiplication. + + Parameters: + - topk_ids: A tensor of shape [total_tokens, top_k] representing the + top-k expert indices for each token. + - block_size: The block size used in block matrix multiplication. + - num_experts: The total number of experts. + - expert_map: A tensor of shape [num_experts] that maps the expert index + from the global space to the local index space of the current + expert parallel shard. If the expert is not in the current expert + parallel shard, the mapping is set to -1. + - pad_sorted_ids: A flag indicating whether the sorted_token_ids length + should be padded to a multiple of block_size, + + Returns: + - sorted_token_ids: A tensor containing the sorted token indices according + to their allocated expert. + - expert_ids: A tensor indicating the assigned expert index for each block. + - num_tokens_post_padded: The total number of tokens after padding, + ensuring divisibility by block_size. + + This function pads the number of tokens that each expert needs to process + so that it is divisible by block_size. + Padding ensures that during block matrix multiplication, the dimensions + align correctly. + + Example: + Given topk_ids = [[2, 3, 4], [1, 2, 4], [1, 3, 4], [1, 2, 3]], + block_size = 4, and num_experts = 4: + - We initially have 12 tokens (after repeating 'top_k' times) and 4 experts, + with each expert needing to process 3 tokens. + - As block_size is 4, we pad 1 token for each expert. + - First, flatten topk_ids to [2, 3, 4, 1, 2, 4, 1, 3, 4, 1, 2, 3]. + - Then append padding tokens [12, 12, 12, 12] for each block. + - After sorting by expert index, we obtain token_ids + [3, 6, 9, 12, 0, 4, 10, 12, 1, 7, 11, 12, 2, 5, 8, 12]. + Tokens 12 are non-existent (padding) and are ignored in + the subsequent matrix multiplication. + - The padding ensures that the total number of tokens is now divisible + by block_size for proper block matrix operations. + """ + max_num_tokens_padded = topk_ids.numel() + num_experts * (block_size - 1) + if pad_sorted_ids: + max_num_tokens_padded = round_up(max_num_tokens_padded, block_size) + sorted_ids = torch.empty( + (max_num_tokens_padded,), dtype=torch.int32, device=topk_ids.device + ) + sorted_ids.fill_(topk_ids.numel()) + max_num_m_blocks = triton.cdiv(max_num_tokens_padded, block_size) + # Expert ids must be zeroed out to prevent index out of bounds error while + # mapping global expert ids to local expert ids in expert parallelism. + expert_ids = torch.zeros( + (max_num_m_blocks,), dtype=torch.int32, device=topk_ids.device + ) + num_tokens_post_pad = torch.empty((1), dtype=torch.int32, device=topk_ids.device) + if num_experts >= 224: + if VLLM_ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON or num_experts != 256: + moe_align_block_size_triton( + topk_ids, + num_experts, + block_size, + sorted_ids, + expert_ids, + num_tokens_post_pad, + ) + else: + # Currently requires num_experts=256 + ops.sgl_moe_align_block_size( + topk_ids, + num_experts, + block_size, + sorted_ids, + expert_ids, + num_tokens_post_pad, + ) + else: + ops.moe_align_block_size( + topk_ids, + num_experts, + block_size, + sorted_ids, + expert_ids, + num_tokens_post_pad, + ) + if expert_map is not None: + expert_ids = expert_map[expert_ids] + + return sorted_ids, expert_ids, num_tokens_post_pad diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/platforms.py b/build/torch26-cxx11-cu126-aarch64-linux/moe/platforms.py new file mode 100644 index 0000000000000000000000000000000000000000..db83d0f2b0fadb15c14e8887ac27e4a07b6c4d62 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/platforms.py @@ -0,0 +1,99 @@ +import random +from abc import ABC, abstractmethod +from functools import lru_cache +from typing import NamedTuple, Optional + +import torch + +IS_ROCM = torch.version.hip is not None + + +class DeviceCapability(NamedTuple): + major: int + minor: int + + def as_version_str(self) -> str: + return f"{self.major}.{self.minor}" + + def to_int(self) -> int: + """ + Express device capability as an integer ````. + + It is assumed that the minor version is always a single digit. + """ + assert 0 <= self.minor < 10 + return self.major * 10 + self.minor + + +class Platform(ABC): + simple_compile_backend: str = "inductor" + + @classmethod + @abstractmethod + def get_device_name(cls, device_id: int = 0) -> str: ... + + @abstractmethod + def is_cuda_alike(self) -> bool: + """Stateless version of :func:`torch.cuda.is_available`.""" + ... + + @abstractmethod + def is_rocm(self): ... + + @classmethod + def seed_everything(cls, seed: Optional[int] = None) -> None: + """ + Set the seed of each random module. + `torch.manual_seed` will set seed on all devices. + + Loosely based on: https://github.com/Lightning-AI/pytorch-lightning/blob/2.4.0/src/lightning/fabric/utilities/seed.py#L20 + """ + if seed is not None: + random.seed(seed) + torch.manual_seed(seed) + try: + import numpy as np + + np.random.seed(seed) + except ImportError: + pass + + +class CudaPlatform(Platform): + @classmethod + def get_device_capability(cls, device_id: int = 0) -> DeviceCapability: + major, minor = torch.cuda.get_device_capability(device_id) + return DeviceCapability(major=major, minor=minor) + + @classmethod + @lru_cache(maxsize=8) + def get_device_name(cls, device_id: int = 0) -> str: + return torch.cuda.get_device_name(0) + + def is_cuda_alike(self) -> bool: + return True + + def is_rocm(self): + return False + + +class RocmPlatform(Platform): + @classmethod + @lru_cache(maxsize=8) + def get_device_capability(cls, device_id: int = 0) -> DeviceCapability: + major, minor = torch.cuda.get_device_capability(device_id) + return DeviceCapability(major=major, minor=minor) + + @classmethod + @lru_cache(maxsize=8) + def get_device_name(cls, device_id: int = 0) -> str: + return torch.cuda.get_device_name(device_id) + + def is_cuda_alike(self) -> bool: + return True + + def is_rocm(self): + return True + + +current_platform = RocmPlatform() if IS_ROCM else CudaPlatform() diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/scalar_type.py b/build/torch26-cxx11-cu126-aarch64-linux/moe/scalar_type.py new file mode 100644 index 0000000000000000000000000000000000000000..ea749fe8247b6846620ccbba30ddf48d914ca4e1 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/scalar_type.py @@ -0,0 +1,336 @@ +import functools +import struct +from dataclasses import dataclass +from enum import Enum +from typing import Optional, Union + + +# Mirrors enum in `core/scalar_type.hpp` +class NanRepr(Enum): + NONE = 0 # nans are not supported + IEEE_754 = 1 # nans are: Exp all 1s, mantissa not all 0s + EXTD_RANGE_MAX_MIN = 2 # nans are: Exp all 1s, mantissa all 1s + + +# This ScalarType class is a parallel implementation of the C++ ScalarType +# class found in csrc/core/scalar_type.hpp. These two classes should be kept +# in sync until the inductor fully supports custom C++ classes. +@dataclass(frozen=True) +class ScalarType: + """ + ScalarType can represent a wide range of floating point and integer + types, in particular it can be used to represent sub-byte data types + (something that torch.dtype currently does not support). It is also + capable of representing types with a bias, i.e.: + `stored_value = value + bias`, + this is useful for quantized types (e.g. standard GPTQ 4bit uses a bias + of 8). The implementation for this class can be found in + csrc/core/scalar_type.hpp, these type signatures should be kept in sync + with that file. + """ + + exponent: int + """ + Number of bits in the exponent if this is a floating point type + (zero if this an integer type) + """ + + mantissa: int + """ + Number of bits in the mantissa if this is a floating point type, + or the number bits representing an integer excluding the sign bit if + this an integer type. + """ + + signed: bool + "If the type is signed (i.e. has a sign bit)" + + bias: int + """ + bias used to encode the values in this scalar type + (value = stored_value - bias, default 0) for example if we store the + type as an unsigned integer with a bias of 128 then the value 0 will be + stored as 128 and -1 will be stored as 127 and 1 will be stored as 129. + """ + + _finite_values_only: bool = False + """ + Private: if infs are supported, used `has_infs()` instead. + """ + + nan_repr: NanRepr = NanRepr.IEEE_754 + """ + How NaNs are represent in this scalar type, returns NanRepr value. + (not applicable for integer types) + """ + + def _floating_point_max_int(self) -> int: + assert ( + self.mantissa <= 52 and self.exponent <= 11 + ), f"Cannot represent max/min as a double for type {self.__str__()}" + + max_mantissa = (1 << self.mantissa) - 1 + if self.nan_repr == NanRepr.EXTD_RANGE_MAX_MIN: + max_mantissa = max_mantissa - 1 + + max_exponent = (1 << self.exponent) - 2 + if self.nan_repr == NanRepr.EXTD_RANGE_MAX_MIN or self.nan_repr == NanRepr.NONE: + assert ( + self.exponent < 11 + ), f"Cannot represent max/min as a double for type {self.__str__()}" + max_exponent = max_exponent + 1 + + # adjust the exponent to match that of a double + # for now we assume the exponent bias is the standard 2^(e-1) -1, (where + # e is the exponent bits), there is some precedent for non-standard + # biases, example `float8_e4m3b11fnuz` here: + # https://github.com/jax-ml/ml_dtypes but to avoid premature over + # complication we are just assuming the standard exponent bias until + # there is a need to support non-standard biases + exponent_bias = (1 << (self.exponent - 1)) - 1 + exponent_bias_double = (1 << 10) - 1 # double e = 11 + + max_exponent_double = max_exponent - exponent_bias + exponent_bias_double + + # shift the mantissa and exponent into the proper positions for an + # IEEE double and bitwise-or them together. + return (max_mantissa << (52 - self.mantissa)) | (max_exponent_double << 52) + + def _floating_point_max(self) -> float: + double_raw = self._floating_point_max_int() + return struct.unpack("!d", struct.pack("!Q", double_raw))[0] + + def _raw_max(self) -> Union[int, float]: + if self.is_floating_point(): + return self._floating_point_max() + else: + assert ( + self.size_bits < 64 or self.size_bits == 64 and self.is_signed() + ), "Cannot represent max as an int" + return (1 << self.mantissa) - 1 + + def _raw_min(self) -> Union[int, float]: + if self.is_floating_point(): + assert ( + self.is_signed() + ), "We currently assume all floating point types are signed" + sign_bit_double = 1 << 63 + + max_raw = self._floating_point_max_int() + min_raw = max_raw | sign_bit_double + return struct.unpack("!d", struct.pack("!Q", min_raw))[0] + else: + assert ( + not self.is_signed() or self.size_bits <= 64 + ), "Cannot represent min as a int64_t" + + if self.is_signed(): + return -(1 << (self.size_bits - 1)) + else: + return 0 + + @functools.cached_property + def id(self) -> int: + """ + Convert the ScalarType to an int which can be passed to pytorch custom + ops. This layout of the int must be kept in sync with the C++ + ScalarType's from_id method. + """ + val = 0 + offset = 0 + + def or_and_advance(member, bit_width): + nonlocal val + nonlocal offset + bit_mask = (1 << bit_width) - 1 + val = val | (int(member) & bit_mask) << offset + offset = offset + bit_width + + or_and_advance(self.exponent, 8) + or_and_advance(self.mantissa, 8) + or_and_advance(self.signed, 1) + or_and_advance(self.bias, 32) + or_and_advance(self._finite_values_only, 1) + or_and_advance(self.nan_repr.value, 8) + + assert offset <= 64, f"ScalarType fields too big {offset} to fit into an int64" + + return val + + @property + def size_bits(self) -> int: + return self.exponent + self.mantissa + int(self.signed) + + def min(self) -> Union[int, float]: + """ + Min representable value for this scalar type. + (accounting for bias if there is one) + """ + return self._raw_min() - self.bias + + def max(self) -> Union[int, float]: + """ + Max representable value for this scalar type. + (accounting for bias if there is one) + """ + return self._raw_max() - self.bias + + def is_signed(self) -> bool: + """ + If the type is signed (i.e. has a sign bit), same as `signed` + added for consistency with: + https://pytorch.org/docs/stable/generated/torch.Tensor.is_signed.html + """ + return self.signed + + def is_floating_point(self) -> bool: + "If the type is a floating point type" + return self.exponent != 0 + + def is_integer(self) -> bool: + "If the type is an integer type" + return self.exponent == 0 + + def has_bias(self) -> bool: + "If the type has a non-zero bias" + return self.bias != 0 + + def has_infs(self) -> bool: + "If the type is floating point and supports infinity" + return not self._finite_values_only + + def has_nans(self) -> bool: + return self.nan_repr != NanRepr.NONE.value + + def is_ieee_754(self) -> bool: + """ + If the type is a floating point type that follows IEEE 754 + conventions + """ + return self.nan_repr == NanRepr.IEEE_754.value and not self._finite_values_only + + def __str__(self) -> str: + """ + naming generally follows: https://github.com/jax-ml/ml_dtypes + for floating point types (leading f) the scheme is: + `float_em[flags]` + flags: + - no-flags: means it follows IEEE 754 conventions + - f: means finite values only (no infinities) + - n: means nans are supported (non-standard encoding) + for integer types the scheme is: + `[u]int[b]` + - if bias is not present it means its zero + """ + if self.is_floating_point(): + ret = ( + "float" + + str(self.size_bits) + + "_e" + + str(self.exponent) + + "m" + + str(self.mantissa) + ) + + if not self.is_ieee_754(): + if self._finite_values_only: + ret = ret + "f" + if self.nan_repr != NanRepr.NONE: + ret = ret + "n" + + return ret + else: + ret = ("int" if self.is_signed() else "uint") + str(self.size_bits) + if self.has_bias(): + ret = ret + "b" + str(self.bias) + return ret + + def __repr__(self) -> str: + return "ScalarType." + self.__str__() + + # __len__ needs to be defined (and has to throw TypeError) for pytorch's + # opcheck to work. + def __len__(self) -> int: + raise TypeError + + # + # Convenience Constructors + # + + @classmethod + def int_(cls, size_bits: int, bias: Optional[int]) -> "ScalarType": + "Create a signed integer scalar type (size_bits includes sign-bit)." + ret = cls(0, size_bits - 1, True, bias if bias else 0) + ret.id # noqa B018: make sure the id is cached + return ret + + @classmethod + def uint(cls, size_bits: int, bias: Optional[int]) -> "ScalarType": + """Create a unsigned integer scalar type.""" + ret = cls(0, size_bits, False, bias if bias else 0) + ret.id # noqa B018: make sure the id is cached + return ret + + @classmethod + def float_IEEE754(cls, exponent: int, mantissa: int) -> "ScalarType": + """ + Create a standard floating point type + (i.e. follows IEEE 754 conventions). + """ + assert mantissa > 0 and exponent > 0 + ret = cls(exponent, mantissa, True, 0) + ret.id # noqa B018: make sure the id is cached + return ret + + @classmethod + def float_( + cls, exponent: int, mantissa: int, finite_values_only: bool, nan_repr: NanRepr + ) -> "ScalarType": + """ + Create a non-standard floating point type + (i.e. does not follow IEEE 754 conventions). + """ + assert mantissa > 0 and exponent > 0 + assert nan_repr != NanRepr.IEEE_754, ( + "use `float_IEEE754` constructor for floating point types that " + "follow IEEE 754 conventions" + ) + ret = cls(exponent, mantissa, True, 0, finite_values_only, nan_repr) + ret.id # noqa B018: make sure the id is cached + return ret + + +# naming generally follows: https://github.com/jax-ml/ml_dtypes +# for floating point types (leading f) the scheme is: +# `float_em[flags]` +# flags: +# - no-flags: means it follows IEEE 754 conventions +# - f: means finite values only (no infinities) +# - n: means nans are supported (non-standard encoding) +# for integer types the scheme is: +# `[u]int[b]` +# - if bias is not present it means its zero + + +class scalar_types: + int4 = ScalarType.int_(4, None) + uint4 = ScalarType.uint(4, None) + int8 = ScalarType.int_(8, None) + uint8 = ScalarType.uint(8, None) + float8_e4m3fn = ScalarType.float_(4, 3, True, NanRepr.EXTD_RANGE_MAX_MIN) + float8_e5m2 = ScalarType.float_IEEE754(5, 2) + float16_e8m7 = ScalarType.float_IEEE754(8, 7) + float16_e5m10 = ScalarType.float_IEEE754(5, 10) + + # fp6, https://github.com/usyd-fsalab/fp6_llm/tree/main + float6_e3m2f = ScalarType.float_(3, 2, True, NanRepr.NONE) + + # "gptq" types + uint2b2 = ScalarType.uint(2, 2) + uint3b4 = ScalarType.uint(3, 4) + uint4b8 = ScalarType.uint(4, 8) + uint8b128 = ScalarType.uint(8, 128) + + # colloquial names + bfloat16 = float16_e8m7 + float16 = float16_e5m10 diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/utils/__init__.py b/build/torch26-cxx11-cu126-aarch64-linux/moe/utils/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/utils/custom_ops.py b/build/torch26-cxx11-cu126-aarch64-linux/moe/utils/custom_ops.py new file mode 100644 index 0000000000000000000000000000000000000000..26ef7c23d1bea7a162d81746ac041baefb472f46 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/utils/custom_ops.py @@ -0,0 +1,68 @@ +from typing import Callable, Optional + +import torch +from torch.library import Library + +from .._ops import ops +from ..platforms import current_platform + + +# Some backends use pytorch version < 2.4.0 which doesn't +# support `torch.library.custom_op`. +def supports_custom_op() -> bool: + return hasattr(torch.library, "custom_op") + + +# TODO: we probably need a shorthand for the ops name. +vllm_lib = Library(ops.__name__.split(".")[-1], "FRAGMENT") + + +def direct_register_custom_op( + op_name: str, + op_func: Callable, + mutates_args: list[str], + fake_impl: Optional[Callable] = None, + target_lib: Optional[Library] = None, + dispatch_key: str = "CUDA", +): + """ + `torch.library.custom_op` can have significant overhead because it + needs to consider complicated dispatching logic. This function + directly registers a custom op and dispatches it to the CUDA backend. + See https://gist.github.com/youkaichao/ecbea9ec9fc79a45d2adce1784d7a9a5 + for more details. + + By default, the custom op is registered to the vLLM library. If you + want to register it to a different library, you can pass the library + object to the `target_lib` argument. + + IMPORTANT: the lifetime of the operator is tied to the lifetime of the + library object. If you want to bind the operator to a different library, + make sure the library object is alive when the operator is used. + """ + if not supports_custom_op(): + from ..platforms import current_platform + + assert not current_platform.is_cuda_alike(), ( + "cuda platform needs torch>=2.4 to support custom op, " + "chances are you are using an old version of pytorch " + "or a custom build of pytorch. It is recommended to " + "use vLLM in a fresh new environment and let it install " + "the required dependencies." + ) + return + + import torch.library + + if hasattr(torch.library, "infer_schema"): + schema_str = torch.library.infer_schema(op_func, mutates_args=mutates_args) + else: + # for pytorch 2.4 + import torch._custom_op.impl + + schema_str = torch._custom_op.impl.infer_schema(op_func, mutates_args) + my_lib = target_lib or vllm_lib + my_lib.define(op_name + schema_str) + my_lib.impl(op_name, op_func, dispatch_key=dispatch_key) + if fake_impl is not None: + my_lib._register_fake(op_name, fake_impl) diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/utils/marlin_utils.py b/build/torch26-cxx11-cu126-aarch64-linux/moe/utils/marlin_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..f27b71af67a6f5f18c8dba271926e3d40d518879 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/utils/marlin_utils.py @@ -0,0 +1,306 @@ +from typing import List, Optional, Tuple + +import numpy +import torch + +from ..scalar_type import ScalarType, scalar_types +from .quant_utils import pack_cols, unpack_cols + +GPTQ_MARLIN_TILE = 16 +GPTQ_MARLIN_MIN_THREAD_N = 64 +GPTQ_MARLIN_MIN_THREAD_K = 128 +GPTQ_MARLIN_MAX_PARALLEL = 16 + +GPTQ_MARLIN_24_TILE = 16 +GPTQ_MARLIN_24_MIN_THREAD_N = 128 +GPTQ_MARLIN_24_MIN_THREAD_K = 128 +GPTQ_MARLIN_24_MAX_PARALLEL = 64 + +GPTQ_MARLIN_24_SUPPORTED_QUANT_TYPES = [scalar_types.uint4b8, scalar_types.uint8b128] +GPTQ_MARLIN_24_SUPPORTED_GROUP_SIZES = [-1, 128] + +MARLIN_QQQ_TILE = 16 +MARLIN_QQQ_MIN_THREAD_N = 64 +MARLIN_QQQ_MIN_THREAD_K = 128 +MARLIN_QQQ_MAX_PARALLEL = 16 + +MARLIN_QQQ_SUPPORTED_NUM_BITS = [4] +MARLIN_QQQ_SUPPORTED_GROUP_SIZES = [-1, 128] +MARLIN_QQQ_SUPPORTED_SYM = [True] + +MARLIN_SUPPORTED_GROUP_SIZES = [-1, 32, 64, 128] + +# In case there is a performance issue with Marlin, the variable below can be +# changed to False, which allows Marlin to perform global reductions in fp16 +# precision (instead of fp32), and therefore, save on some memory movements. +USE_FP32_REDUCE_DEFAULT = True + + +# For binary size and compile time, we don't support the same types for with and +# without runtime zero-point. We support common cases, i.e. AWQ and GPTQ. +# TODO: we may want to move this into the C++ so its closer to the actual impl +def query_marlin_supported_quant_types( + has_zp: bool, device_capability: Optional[int] = None +): + if device_capability is None: + capability_tuple = torch.cuda.get_device_capability() + device_capability = capability_tuple[0] * 10 + capability_tuple[1] + + if device_capability < 80: + return [] + + if has_zp: + # AWQ style, unsigned + runtime zero-point + return [scalar_types.uint4, scalar_types.uint8] + else: + # GPTQ style, unsigned + symmetric bias + # TODO: once fp8_marlin is merged into "gptq_marlin" we should be able + # to add `scalar_types.float8_e4m3fn` here + return [scalar_types.uint4b8, scalar_types.uint8b128] + + +def _check_marlin_supported( + quant_type: ScalarType, + group_size: Optional[int], + has_zp: bool, + device_capability: Optional[int] = None, +) -> Tuple[bool, Optional[str]]: + + if device_capability is None: + capability_tuple = torch.cuda.get_device_capability() + device_capability = capability_tuple[0] * 10 + capability_tuple[1] + + supported_types = query_marlin_supported_quant_types(has_zp, device_capability) + + if quant_type not in supported_types: + return ( + False, + f"Marlin does not support weight_bits = {quant_type}. " + f"Only types = {supported_types} " + f"are supported (for group_size = {group_size}, " + f"device_capability = {device_capability}, zp = {has_zp}).", + ) + if group_size is None or group_size not in MARLIN_SUPPORTED_GROUP_SIZES: + return ( + False, + f"Marlin does not support group_size = {group_size}. " + f"Only group_sizes = {MARLIN_SUPPORTED_GROUP_SIZES} " + "are supported.", + ) + + return True, None + + +def check_marlin_supported( + quant_type: ScalarType, + group_size: int, + has_zp: bool = False, + device_capability: Optional[int] = None, +) -> bool: + cond, _ = _check_marlin_supported(quant_type, group_size, has_zp, device_capability) + return cond + + +def verify_marlin_supported( + quant_type: ScalarType, group_size: int, has_zp: bool = False +) -> None: + cond, err_msg = _check_marlin_supported(quant_type, group_size, has_zp) + if not cond: + assert err_msg is not None + raise ValueError(err_msg) + + +def verify_marlin_supports_shape( + output_size_per_partition: int, + input_size_per_partition: int, + input_size: int, + group_size: int, +) -> None: + + # Validate output_size_per_partition + if output_size_per_partition % GPTQ_MARLIN_MIN_THREAD_N != 0: + raise ValueError( + f"Weight output_size_per_partition = " + f"{output_size_per_partition} is not divisible by " + f" min_thread_n = {GPTQ_MARLIN_MIN_THREAD_N}. " + "Consider reducing tensor_parallel_size or running " + "with --quantization gptq." + ) + + # Validate input_size_per_partition + if input_size_per_partition % GPTQ_MARLIN_MIN_THREAD_K != 0: + raise ValueError( + f"Weight input_size_per_partition = " + f"{input_size_per_partition} is not divisible " + f"by min_thread_k = {GPTQ_MARLIN_MIN_THREAD_K}. " + "Consider reducing tensor_parallel_size or running " + "with --quantization gptq." + ) + + if group_size < input_size and input_size_per_partition % group_size != 0: + raise ValueError( + f"Weight input_size_per_partition = {input_size_per_partition}" + f" is not divisible by group_size = {group_size}." + "Consider reducing tensor_parallel_size or running " + "with --quantization gptq." + ) + + +def check_marlin_supports_shape( + output_size_per_partition: int, + input_size_per_partition: int, + input_size: int, + group_size: int, +) -> Tuple[bool, Optional[str]]: + try: + verify_marlin_supports_shape( + output_size_per_partition, input_size_per_partition, input_size, group_size + ) + except ValueError as e: + return False, e.__str__() + return True, None + + +def marlin_make_workspace( + output_size_per_partition: int, device: torch.device +) -> torch.Tensor: + max_workspace_size = ( + output_size_per_partition // GPTQ_MARLIN_MIN_THREAD_N + ) * GPTQ_MARLIN_MAX_PARALLEL + + return torch.zeros( + max_workspace_size, dtype=torch.int, device=device, requires_grad=False + ) + + +def marlin_is_k_full(act_order: bool, is_row_parallel: bool) -> bool: + return (not act_order) or (act_order and not is_row_parallel) + + +def marlin_repeat_scales_on_all_ranks( + act_order: bool, group_size: int, is_row_parallel: bool +) -> bool: + # Need to repeat scales on every rank if act_ordering or + # channelwise and RowParallelLinear + is_channelwise = group_size == -1 + return act_order or (is_channelwise and is_row_parallel) + + +def marlin_make_empty_g_idx(device: torch.device) -> torch.Tensor: + return torch.nn.Parameter( + torch.empty(0, dtype=torch.int, device=device), requires_grad=False + ) + + +def marlin_make_empty_zp(device: torch.device) -> torch.Tensor: + return torch.nn.Parameter( + torch.empty(0, dtype=torch.int, device=device), requires_grad=False + ) + + +def marlin_sort_g_idx(g_idx: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]: + g_idx_sort_indices = torch.argsort(g_idx).to(torch.int) + return g_idx[g_idx_sort_indices], g_idx_sort_indices + + +def get_scale_perms(): + scale_perm: List[int] = [] + for i in range(8): + scale_perm.extend([i + 8 * j for j in range(8)]) + scale_perm_single: List[int] = [] + for i in range(4): + scale_perm_single.extend([2 * i + j for j in [0, 1, 8, 9, 16, 17, 24, 25]]) + return scale_perm, scale_perm_single + + +def marlin_permute_scales( + s: torch.Tensor, size_k: int, size_n: int, group_size: int +) -> torch.Tensor: + + scale_perm, scale_perm_single = get_scale_perms() + if group_size < size_k and group_size != -1: + s = s.reshape((-1, len(scale_perm)))[:, scale_perm] + else: + s = s.reshape((-1, len(scale_perm_single)))[:, scale_perm_single] + s = s.reshape((-1, size_n)).contiguous() + + return s + + +def marlin_moe_permute_scales( + s: torch.Tensor, + size_k: int, + size_n: int, + group_size: int, +): + num_experts = s.shape[0] + output = torch.empty( + (num_experts, s.shape[1], s.shape[2]), + device=s.device, + dtype=s.dtype, + ) + + for e in range(num_experts): + output[e] = marlin_permute_scales(s[e], size_k, size_n, group_size) + return output + + +def marlin_zero_points( + zp: torch.Tensor, size_k: int, size_n: int, num_bits: int +) -> torch.Tensor: + # Permute zero-points in a similar way to scales, but do not use the + # "single" permutation, since zero-points are applied on every MMA + scale_perm, _ = get_scale_perms() + zp = zp.reshape((-1, len(scale_perm)))[:, scale_perm] + + # Interleave column dim (for the dequantize code) and pack it to int32 + if num_bits == 4: + interleave = numpy.array([0, 2, 4, 6, 1, 3, 5, 7]) + elif num_bits == 8: + interleave = numpy.array([0, 2, 1, 3]) + else: + raise Exception("num_bits must be 4 or 8, got {}".format(num_bits)) + + zp = zp.reshape((-1, len(interleave)))[:, interleave].ravel() + zp = zp.reshape((-1, size_n)).contiguous() + zp = pack_cols(zp, num_bits, size_k, size_n) + + return zp + + +def awq_to_marlin_zero_points( + q_zp_packed: torch.Tensor, size_k: int, size_n: int, num_bits: int +) -> torch.Tensor: + # AWQ zero-points are quantized and packed on the column dim. + # In addition, the values are permuted based on dequantizer. + # Here we undo both of these, and then apply marlin permutation + # and pack it back. + q_zp = unpack_cols(q_zp_packed, num_bits, size_k, size_n) + + # Undo interleaving (use argsort(..) to get inverse perm) + if num_bits == 4: + undo_interleave = numpy.argsort(numpy.array([0, 2, 4, 6, 1, 3, 5, 7])) + elif num_bits == 8: + undo_interleave = numpy.argsort(numpy.array([0, 2, 1, 3])) + else: + raise Exception("num_bits must be 4 or 8, got {}".format(num_bits)) + + q_zp = q_zp.reshape((-1, len(undo_interleave)))[:, undo_interleave].ravel() + q_zp = q_zp.reshape((-1, size_n)).contiguous() + + marlin_zp = marlin_zero_points(q_zp, size_k, size_n, num_bits) + return marlin_zp + + +def moe_awq_to_marlin_zero_points( + q_zp_packed: torch.Tensor, size_k: int, size_n: int, num_bits: int +): + num_experts = q_zp_packed.shape[0] + output = torch.empty( + (num_experts, q_zp_packed.shape[1], q_zp_packed.shape[2]), + device=q_zp_packed.device, + dtype=q_zp_packed.dtype, + ) + for e in range(num_experts): + output[e] = awq_to_marlin_zero_points(q_zp_packed[e], size_k, size_n, num_bits) + return output diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/utils/marlin_utils_test.py b/build/torch26-cxx11-cu126-aarch64-linux/moe/utils/marlin_utils_test.py new file mode 100644 index 0000000000000000000000000000000000000000..b03555f5bffa8d48187d37e53f17e6e3138472cc --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/utils/marlin_utils_test.py @@ -0,0 +1,161 @@ +"""Utility functions used for tests and benchmarks""" + +from typing import List, Optional + +import numpy as np +import torch + +from ..scalar_type import ScalarType +from .marlin_utils import GPTQ_MARLIN_TILE, marlin_permute_scales, marlin_zero_points +from .quant_utils import ( + get_pack_factor, + gptq_quantize_weights, + quantize_weights, + sort_weights, +) + + +class MarlinWorkspace: + + def __init__(self, out_features, min_thread_n, max_parallel): + assert ( + out_features % min_thread_n == 0 + ), "out_features = {} is undivisible by min_thread_n = {}".format( + out_features, min_thread_n + ) + + max_workspace_size = (out_features // min_thread_n) * max_parallel + + self.scratch = torch.zeros(max_workspace_size, dtype=torch.int, device="cuda") + + +def marlin_permute_weights(q_w, size_k, size_n, perm, tile=GPTQ_MARLIN_TILE): + assert q_w.shape == (size_k, size_n) + assert size_k % tile == 0, f"size_k = {size_k}, tile = {tile}" + assert size_n % tile == 0, f"size_k = {size_n}, tile = {tile}" + + # Permute weights to 16x64 marlin tiles + q_w = q_w.reshape((size_k // tile, tile, size_n // tile, tile)) + q_w = q_w.permute((0, 2, 1, 3)) + q_w = q_w.reshape((size_k // tile, size_n * tile)) + + q_w = q_w.reshape((-1, perm.numel()))[:, perm].reshape(q_w.shape) + + return q_w + + +def marlin_weights(q_w, size_k, size_n, num_bits, perm): + # Permute + q_w = marlin_permute_weights(q_w, size_k, size_n, perm) + + # Pack + pack_factor = get_pack_factor(num_bits) + orig_device = q_w.device + + q_w = q_w.cpu().numpy().astype(np.uint32) + + q_packed = np.zeros((q_w.shape[0], q_w.shape[1] // pack_factor), dtype=np.uint32) + for i in range(pack_factor): + q_packed |= q_w[:, i::pack_factor] << num_bits * i + + q_packed = torch.from_numpy(q_packed.astype(np.int32)).to(orig_device) + + return q_packed + + +def get_weight_perm(num_bits: int): + perm_list: List[int] = [] + for i in range(32): + perm1: List[int] = [] + col = i // 4 + for block in [0, 1]: + for row in [ + 2 * (i % 4), + 2 * (i % 4) + 1, + 2 * (i % 4 + 4), + 2 * (i % 4 + 4) + 1, + ]: + perm1.append(16 * row + col + 8 * block) + for j in range(4): + perm_list.extend([p + 256 * j for p in perm1]) + + perm = np.array(perm_list) + + if num_bits == 4: + interleave = np.array([0, 2, 4, 6, 1, 3, 5, 7]) + elif num_bits == 8: + interleave = np.array([0, 2, 1, 3]) + else: + raise Exception("num_bits must be 4 or 8, got {}".format(num_bits)) + + perm = perm.reshape((-1, len(interleave)))[:, interleave].ravel() + perm = torch.from_numpy(perm) + return perm + + +def marlin_quantize( + w: torch.Tensor, + quant_type: ScalarType, + group_size: int, + act_order: bool, + test_perm: Optional[torch.Tensor] = None, +): + size_k, size_n = w.shape + num_bits = quant_type.size_bits + + # Normalize group_size + if group_size == -1: + group_size = size_k + assert group_size <= size_k + + # Quantize (and apply act_order if provided) + w_ref, q_w, s, g_idx, rand_perm = gptq_quantize_weights( + w, quant_type, group_size, act_order, test_perm + ) + + # For act_order, sort the "weights" and "g_idx" so that group ids are + # increasing + sort_indices = torch.empty(0, dtype=torch.int, device=w.device) + if act_order: + q_w, g_idx, sort_indices = sort_weights(q_w, g_idx) + + # Reformat to marlin + weight_perm = get_weight_perm(num_bits) + marlin_q_w = marlin_weights(q_w, size_k, size_n, num_bits, weight_perm) + marlin_s = marlin_permute_scales(s, size_k, size_n, group_size) + + # Create result + res_list = [w_ref, marlin_q_w, marlin_s, g_idx, sort_indices, rand_perm] + for i in range(len(res_list)): + res_list[i] = res_list[i].to(w.device) + + return res_list + + +def awq_marlin_quantize(w: torch.Tensor, quant_type: ScalarType, group_size: int): + size_k, size_n = w.shape + + # Normalize group_size + if group_size == -1: + group_size = size_k + assert group_size <= size_k + + # Detect num groups + assert size_k % group_size == 0 + num_groups = size_k // group_size + + # Quantize with zp + w_ref, q_w, s, zp = quantize_weights(w, quant_type, group_size, zero_points=True) + + # Reformat to marlin + weight_perm = get_weight_perm(quant_type.size_bits) + marlin_q_w = marlin_weights(q_w, size_k, size_n, quant_type.size_bits, weight_perm) + marlin_s = marlin_permute_scales(s, size_k, size_n, group_size) + marlin_zp = marlin_zero_points(zp, num_groups, size_n, quant_type.size_bits) + + # Create result + res_list = [w_ref, marlin_q_w, marlin_s, marlin_zp] + for i in range(len(res_list)): + res_list[i] = res_list[i].to(w.device) + + return res_list diff --git a/build/torch26-cxx11-cu126-aarch64-linux/moe/utils/quant_utils.py b/build/torch26-cxx11-cu126-aarch64-linux/moe/utils/quant_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..65c90821773b60cca2b0f9102ca7ad4e3cd009d5 --- /dev/null +++ b/build/torch26-cxx11-cu126-aarch64-linux/moe/utils/quant_utils.py @@ -0,0 +1,470 @@ +"""This file is used for /tests and /benchmarks""" + +from typing import List, Optional + +import numpy +import torch + +from ..scalar_type import ScalarType, scalar_types + +SUPPORTED_GPTQ_QUANT_TYPES = [scalar_types.uint4b8, scalar_types.uint8b128] +SUPPORTED_GROUP_SIZES = [-1, 32, 64, 128] + +MARLIN_QQQ_SUPPORTED_NUM_BITS = [4] + +# Note: this is a hack. We should update each model to register the +# stacked params and get it from there instead in a future PR. +# fused_name: List[shard_name] +FUSED_LAYER_NAME_MAPPING = { + "qkv_proj": ["q_proj", "k_proj", "v_proj"], + "gate_up_proj": ["gate_proj", "up_proj"], +} + + +def pack_quantized_values_into_int32( + w_q: torch.Tensor, wtype: ScalarType, packed_dim: int = 0 +): + # move dim to pack to the end + perm = (*[i for i in range(len(w_q.shape)) if i != packed_dim], packed_dim) + inv_perm = tuple(perm.index(i) for i in range(len(perm))) + w_q_perm = w_q.permute(perm) + + pack_factor = 32 // wtype.size_bits + mask = (1 << wtype.size_bits) - 1 + + new_shape_perm = list(w_q_perm.shape) + assert w_q_perm.shape[-1] % pack_factor == 0 + new_shape_perm[-1] //= pack_factor + + res = torch.zeros(new_shape_perm, dtype=torch.int32, device=w_q.device) + for i in range(pack_factor): + res |= (w_q_perm[..., i::pack_factor] & mask) << wtype.size_bits * i + + return res.permute(inv_perm) + + +def unpack_quantized_values_into_int32( + w_q: torch.Tensor, wtype: ScalarType, packed_dim: int = 0 +): + # move dim to pack to the end + perm = (*[i for i in range(len(w_q.shape)) if i != packed_dim], packed_dim) + inv_perm = tuple(perm.index(i) for i in range(len(perm))) + w_q_perm = w_q.permute(perm) + + pack_factor = 32 // wtype.size_bits + mask = (1 << wtype.size_bits) - 1 + + new_shape_perm = list(w_q_perm.shape) + new_shape_perm[-1] *= pack_factor + + res = torch.zeros(new_shape_perm, dtype=torch.int32, device=w_q.device) + for i in range(pack_factor): + res[..., i::pack_factor] = (w_q_perm >> wtype.size_bits * i) & mask + + return res.permute(inv_perm) + + +def is_layer_skipped(prefix: str, ignored_layers: List[str]) -> bool: + # prefix: model.layers.0.self_attn.q_proj + # proj_name: q_proj + proj_name = prefix.split(".")[-1] + if proj_name in FUSED_LAYER_NAME_MAPPING: + shard_prefixes = [ + prefix.replace(proj_name, shard_proj_name) + for shard_proj_name in FUSED_LAYER_NAME_MAPPING[proj_name] + ] + + is_skipped = None + for shard_prefix in shard_prefixes: + is_shard_skipped = shard_prefix in ignored_layers + + if is_skipped is None: + is_skipped = is_shard_skipped + elif is_shard_skipped != is_skipped: + raise ValueError( + f"Detected some but not all shards of {prefix} " + "are quantized. All shards of fused layers " + "to have the same precision." + ) + else: + is_skipped = prefix in ignored_layers + + assert is_skipped is not None + return is_skipped + + +def get_pack_factor(num_bits): + assert 32 % num_bits == 0, f"Unsupported num_bits = {num_bits}" + return 32 // num_bits + + +def permute_rows( + q_w: torch.Tensor, + w_ref: torch.Tensor, + group_size: int, + test_perm: Optional[torch.Tensor] = None, +): + assert q_w.shape == w_ref.shape + + orig_device = q_w.device + k_size, _ = q_w.shape + + g_idx = torch.zeros((k_size,), dtype=torch.int32) + for i in range(k_size): + g_idx[i] = i // group_size + + # Simulate act_order by doing a random permutation on K + rand_perm = test_perm if test_perm is not None else torch.randperm(k_size) + + g_idx = g_idx[rand_perm].contiguous() + q_w = q_w[rand_perm, :].contiguous() + w_ref = w_ref[rand_perm, :].contiguous() + + return ( + w_ref.to(device=orig_device), + q_w.to(device=orig_device), + g_idx.to(device=orig_device), + rand_perm.to(device=orig_device), + ) + + +def quantize_weights( + w: torch.Tensor, + quant_type: ScalarType, + group_size: Optional[int], + zero_points: bool = False, + ref_zero_points_after_scales: bool = False, +): + assert ( + quant_type.is_integer() + ), "Floating point quantization may work but has not been tested" + assert not zero_points or group_size is not None, ( + "to have group zero points, group_size must be provided " + "(-1 group_size is channelwise)" + ) + + orig_device = w.device + orig_type = w.dtype + size_k, size_n = w.shape + + assert w.is_floating_point(), "w must be float" + + if group_size == -1: + group_size = size_k + + # Reshape to [groupsize, -1] + if group_size is not None and group_size < size_k: + w = w.reshape((-1, group_size, size_n)) + w = w.permute(1, 0, 2) + w = w.reshape((group_size, -1)) + + # Compute scale for each group + max_val = torch.max(w, 0, keepdim=True).values + min_val = torch.min(w, 0, keepdim=True).values + + max_q_val = quant_type.max() + min_q_val = quant_type.min() + + w_s = torch.Tensor([1.0]).to(w.device) # unscaled case + maybe_w_zp = None + if group_size is not None: + if zero_points: + assert not quant_type.is_signed() and quant_type.max() > 0 + w_s = (max_val - min_val).clamp(min=1e-5) / quant_type.max() + maybe_w_zp = ( + torch.round(torch.abs(min_val / w_s)).clamp(min_q_val, max_q_val).int() + ) + else: + # If the bias is such that there are no possible negative/positive + # values, set the max value to inf to avoid divide by 0 + w_s = torch.max( + abs(max_val / (max_q_val if max_q_val != 0 else torch.inf)), + abs(min_val / (min_q_val if min_q_val != 0 else torch.inf)), + ) + + # Quantize + w_q = torch.round(w / w_s).int() + (maybe_w_zp if zero_points else 0) + w_q = torch.clamp(w_q, min_q_val, max_q_val) + + # Compute ref (dequantized) + # For some kernels (namely Machete) the zero-points are applied after the + # scales are applied, for this case computing the reference in similar way + # allows us to use tighter error tolerances in our unit tests. + if ref_zero_points_after_scales and maybe_w_zp is not None: + w_ref = w_q.to(orig_type) * w_s - maybe_w_zp.to(orig_type) * w_s + else: + w_ref = (w_q - (maybe_w_zp if zero_points else 0)).to(orig_type) * w_s + + if quant_type.has_bias(): + w_q += quant_type.bias + + # Restore original shapes + if group_size is not None and group_size < size_k: + + def reshape_w(w): + w = w.reshape((group_size, -1, size_n)) + w = w.permute(1, 0, 2) + w = w.reshape((size_k, size_n)).contiguous() + return w + + w_q = reshape_w(w_q) + w_ref = reshape_w(w_ref) + w_s = w_s.reshape((-1, size_n)).contiguous() + + if maybe_w_zp is not None: + maybe_w_zp = maybe_w_zp.reshape((-1, size_n)).contiguous() + maybe_w_zp = maybe_w_zp.to(device=orig_device) + + return ( + w_ref.to(device=orig_device), + w_q.to(device=orig_device), + w_s if group_size is not None else None, + maybe_w_zp, + ) + + +def gptq_quantize_weights( + w: torch.Tensor, + quant_type: ScalarType, + group_size: int, + act_order: bool, + test_perm: Optional[torch.Tensor] = None, +): + size_k, _ = w.shape + + assert w.is_floating_point(), "w must be float" + assert ( + quant_type in SUPPORTED_GPTQ_QUANT_TYPES + ), f"Unsupported gptq type = {quant_type}" + assert group_size in SUPPORTED_GROUP_SIZES + [ + size_k + ], f"Unsupported groupsize = {group_size}" + + w_ref, w_q, w_s, _ = quantize_weights(w, quant_type, group_size) + + # Apply act_order + g_idx = torch.empty(0, dtype=torch.int, device=w.device) + rand_perm = torch.empty(0, dtype=torch.int, device=w.device) + if act_order: + assert ( + group_size < size_k + ), "For act_order, groupsize = {} must be less than size_k = {}".format( + group_size, size_k + ) + + w_ref, w_q, g_idx, rand_perm = permute_rows(w_q, w_ref, group_size, test_perm) + + return w_ref, w_q, w_s, g_idx, rand_perm + + +# QQQ employs different quant schemes for per-group and +# per-channel quantization. +def qqq_quantize_weights(w: torch.Tensor, num_bits: int, group_size: int): + orig_device = w.device + size_k, size_n = w.shape + + assert w.is_floating_point(), "w must be float" + assert ( + num_bits in MARLIN_QQQ_SUPPORTED_NUM_BITS + ), f"Unsupported num_bits = {num_bits}" + assert group_size in SUPPORTED_GROUP_SIZES + [ + size_k + ], f"Unsupported groupsize = {group_size}" + + if group_size == -1: + group_size = size_k + assert group_size <= size_k + + if group_size < size_k: + # Reshape to [groupsize, -1] + w = w.reshape((-1, group_size, size_n)) + w = w.permute(1, 0, 2) + w = w.reshape((group_size, -1)) + + max_q_val = 2**num_bits - 1 + half_q_val = (max_q_val + 1) // 2 + + # Compute scale for each group + s_group = torch.max(torch.abs(w), 0, keepdim=True)[0] + s_group *= 2 / max_q_val # 2 => symmetric + + # Quantize + q_w = torch.round(w / s_group).int() + q_w += half_q_val + q_w = torch.clamp(q_w, 0, max_q_val) + # Compute ref (dequantized) + w_ref = (q_w - half_q_val).half() * s_group + + # Restore original shapes + def reshape_w(w): + w = w.reshape((group_size, -1, size_n)) + w = w.permute(1, 0, 2) + w = w.reshape((size_k, size_n)).contiguous() + return w + + q_w = reshape_w(q_w) + w_ref = reshape_w(w_ref) + + # Compute int8 quantization scale for each channel + s_channel = torch.max(torch.abs(w_ref), 0, keepdim=True)[0] + s_channel /= 127.0 + t_int8 = (w_ref / s_channel).round().clamp(-128, 127).to(torch.int8) + w_ref = t_int8.half() * s_channel + s_channel = s_channel.reshape(1, -1).to(dtype=torch.float) + + # Fuse scales + s_group = (s_group.reshape(-1, size_n).contiguous() / s_channel).to( + dtype=torch.half + ) + else: + max_q_val = 2 ** (num_bits - 1) - 1 + + # Compute scale for each channel + s_channel = torch.max(torch.abs(w), 0, keepdim=True)[0] + s_channel /= max_q_val + + # Quantize + q_w = torch.round(w / s_channel).int() + q_w = torch.clamp(q_w, -max_q_val, max_q_val) + # Compute ref (dequantized) + w_ref = q_w.half() * s_channel + + s_group = torch.tensor([], dtype=torch.half) + # div 2 ** (8 - self.bits)) to offset right shift in unpacking + s_channel /= 2 ** (8 - num_bits) + s_channel = s_channel.reshape(-1, size_n).contiguous().to(torch.float) + + return ( + w_ref.to(device=orig_device), + q_w.to(device=orig_device), + s_group.to(device=orig_device), + s_channel.to(device=orig_device), + ) + + +def sort_weights(q_w: torch.Tensor, g_idx: torch.Tensor): + orig_device = q_w.device + + sort_indices = torch.argsort(g_idx).to(dtype=torch.int32) # Sort based on g_idx + + g_idx = g_idx[sort_indices].contiguous() + q_w = q_w[sort_indices, :].contiguous() + + return ( + q_w.to(device=orig_device), + g_idx.to(device=orig_device), + sort_indices.to(device=orig_device), + ) + + +def pack_rows( + q_w: torch.Tensor, + num_bits: int, + size_k: int, + size_n: int, +): + assert q_w.shape == (size_k, size_n) + + pack_factor = get_pack_factor(num_bits) + assert size_k % pack_factor == 0 + + orig_device = q_w.device + + q_w = q_w.cpu().numpy().astype(numpy.uint32) + + q_res = numpy.zeros((size_k // pack_factor, size_n), dtype=numpy.uint32) + + for i in range(pack_factor): + q_res |= q_w[i::pack_factor, :] << num_bits * i + + q_res = torch.from_numpy(q_res.astype(numpy.int32)).to(orig_device) + return q_res + + +def pack_cols( + q_w: torch.Tensor, + num_bits: int, + size_k: int, + size_n: int, +): + assert q_w.shape == (size_k, size_n) + + pack_factor = get_pack_factor(num_bits) + assert size_n % pack_factor == 0 + + orig_device = q_w.device + + q_w = q_w.cpu().numpy().astype(numpy.uint32) + + q_res = numpy.zeros((size_k, size_n // pack_factor), dtype=numpy.uint32) + + for i in range(pack_factor): + q_res |= q_w[:, i::pack_factor] << num_bits * i + + q_res = torch.from_numpy(q_res.astype(numpy.int32)).to(orig_device) + q_res = q_res.contiguous() + + return q_res + + +def unpack_cols( + packed_q_w: torch.Tensor, + num_bits: int, + size_k: int, + size_n: int, +): + pack_factor = get_pack_factor(num_bits) + assert size_n % pack_factor == 0 + assert packed_q_w.shape == ( + size_k, + size_n // pack_factor, + ), "packed_q_w.shape = {} size_k = {}, size_n = {} pack_Factor = {}".format( + packed_q_w.shape, size_k, size_n, pack_factor + ) + + orig_device = packed_q_w.device + + packed_q_w_cpu = packed_q_w.cpu().numpy().astype(numpy.uint32) + q_res = numpy.zeros((size_k, size_n), dtype=numpy.uint32) + + mask = (1 << num_bits) - 1 + for i in range(pack_factor): + vals = packed_q_w_cpu & mask + packed_q_w_cpu >>= num_bits + q_res[:, i::pack_factor] = vals + + q_res = torch.from_numpy(q_res.astype(numpy.int32)).to(orig_device) + q_res = q_res.contiguous() + + return q_res + + +def gptq_pack( + q_w: torch.Tensor, + num_bits: int, + size_k: int, + size_n: int, +): + return pack_rows(q_w, num_bits, size_k, size_n) + + +def awq_pack( + q_w: torch.Tensor, + num_bits: int, + size_k: int, + size_n: int, +): + assert q_w.shape == (size_k, size_n) + + # Interleave column dim (for the dequantize code) and pack it to int32 + if num_bits == 4: + interleave = numpy.array([0, 2, 4, 6, 1, 3, 5, 7]) + elif num_bits == 8: + interleave = numpy.array([0, 2, 1, 3]) + else: + raise Exception("num_bits must be 4 or 8, got {}".format(num_bits)) + + q_w = q_w.reshape((-1, len(interleave)))[:, interleave].ravel() + q_w = q_w.reshape((-1, size_n)).contiguous() + + return pack_cols(q_w, num_bits, size_k, size_n) diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/__init__.py b/build/torch26-cxx98-cu126-aarch64-linux/moe/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..0d9d8227711def6d51c24b5fd779a3d13681d88c --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/__init__.py @@ -0,0 +1,80 @@ +import torch + +from . import layers +from ._ops import ops +from .fp8_utils import per_token_group_quant_fp8, w8a8_block_fp8_matmul +from .fused_marlin_moe import fused_marlin_moe +from .fused_moe import fused_experts, fused_moe, fused_topk, grouped_topk +from .scalar_type import ScalarType, scalar_types + + +def gptq_marlin_moe_repack( + b_q_weight: torch.Tensor, + perm: torch.Tensor, + size_k: int, + size_n: int, + num_bits: int, +) -> torch.Tensor: + num_experts = b_q_weight.shape[0] + assert size_k % 16 == 0 + output = torch.empty( + (num_experts, size_k // 16, size_n * (num_bits // 2)), + device=b_q_weight.device, + dtype=b_q_weight.dtype, + ) + for e in range(num_experts): + output[e] = ops.gptq_marlin_repack( + b_q_weight[e], perm[e], size_k, size_n, num_bits + ) + return output + + +def awq_marlin_moe_repack( + b_q_weight: torch.Tensor, + perm: torch.Tensor, + size_k: int, + size_n: int, + num_bits: int, +) -> torch.Tensor: + num_experts = b_q_weight.shape[0] + assert size_k % 16 == 0 + output = torch.empty( + (num_experts, size_k // 16, size_n * (num_bits // 2)), + device=b_q_weight.device, + dtype=b_q_weight.dtype, + ) + for e in range(num_experts): + output[e] = ops.awq_marlin_repack(b_q_weight[e], size_k, size_n, num_bits) + return output + + +def moe_sum(input: torch.Tensor, output: torch.Tensor): + ops.moe_sum(input, output) + + +def topk_softmax( + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + token_expert_indicies: torch.Tensor, + gating_output: float, +) -> None: + ops.topk_softmax(topk_weights, topk_ids, token_expert_indicies, gating_output) + + +__all__ = [ + "ScalarType", + "awq_marlin_moe_repack", + "fused_experts", + "fused_marlin_moe", + "fused_moe", + "fused_topk", + "gptq_marlin_moe_repack", + "grouped_topk", + "layers", + "moe_align_block_size", + "moe_sum", + "per_token_group_quant_fp8", + "scalar_types", + "topk_softmax", + "w8a8_block_fp8_matmul", +] diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/_moe_69e0359.abi3.so b/build/torch26-cxx98-cu126-aarch64-linux/moe/_moe_69e0359.abi3.so new file mode 100755 index 0000000000000000000000000000000000000000..13d6550482166b43f942fc15193ab57a7f19a206 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/_moe_69e0359.abi3.so @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f3189cf93c288e1e44af2e3715459cf96185978b1241f233f569153a5861555 +size 67581968 diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/_ops.py b/build/torch26-cxx98-cu126-aarch64-linux/moe/_ops.py new file mode 100644 index 0000000000000000000000000000000000000000..b33dfcc6afacf69f6989c184c5c3efabc03f18da --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/_ops.py @@ -0,0 +1,9 @@ +import torch +from . import _moe_69e0359 +ops = torch.ops._moe_69e0359 + +def add_op_namespace_prefix(op_name: str): + """ + Prefix op by namespace. + """ + return f"_moe_69e0359::{op_name}" \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..56c1a4e3af0b4a93fff71028d8e04bf73f0abb29 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..d3677bebb82a7f3f19344ef6471626493cf2c5bb --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=1,N=1792,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=1,N=1792,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..265768fb900ccfe9612b4a0d25973e6618f22a79 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=1,N=1792,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "5120": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "9216": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "13312": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "17408": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "25600": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "33792": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "41984": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "50176": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "58368": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=1,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=1,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..d3be23dfc903ba61d3d4d79c0230952b24d2ead0 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=1,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "5120": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "9216": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "13312": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "17408": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "25600": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "33792": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "41984": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "50176": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "58368": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=1,N=3072,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=1,N=3072,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..589f5d39f31418d5121e7cbb2e6f2894b0a7ed32 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=1,N=3072,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "5120": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "9216": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "13312": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "17408": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "25600": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "33792": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "41984": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "50176": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "58368": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=1,N=3072,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=1,N=3072,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..2c78bfaba7890772bf266721f5577202ea443882 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=1,N=3072,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "5120": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "9216": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "13312": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "17408": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "25600": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "33792": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "41984": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "50176": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "58368": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=1,N=3072,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=1,N=3072,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..4da841e74a79f9589fecac1fa557ea132d34805f --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=1,N=3072,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "5120": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "9216": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "13312": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "17408": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "25600": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "33792": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "41984": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "50176": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "58368": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=1,N=3584,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=1,N=3584,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..200356713c0d0a76e199671c7ec8f10d0e5ee0ac --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=1,N=3584,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "5120": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "9216": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "13312": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "17408": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "25600": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "33792": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "41984": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "50176": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "58368": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=1,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=1,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..e076615ee541a5043556f630ecf0946c4e2c1408 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=1,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "5120": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "9216": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "13312": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "17408": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "25600": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "33792": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "41984": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "50176": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "58368": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=1,N=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=1,N=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..ee896554b921040d7810bb6e9368cc200777951d --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=1,N=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "5120": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "9216": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "13312": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "17408": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "25600": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "33792": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "41984": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "50176": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "58368": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=1,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=1,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..05aed8b1c81492151d128ef251afc510d8cc8ed5 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=1,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "5120": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "9216": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "13312": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "17408": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "25600": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "33792": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "41984": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "50176": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "58368": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=1344,device_name=NVIDIA_A100-SXM4-40GB.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=1344,device_name=NVIDIA_A100-SXM4-40GB.json new file mode 100644 index 0000000000000000000000000000000000000000..9262a74a4a0e1e3789f260a3ef7f6cb9551f3f2b --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=1344,device_name=NVIDIA_A100-SXM4-40GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=1344,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=1344,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..d251f9b5accaec977fc87a0999cd56ee387fc650 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=1344,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=1344,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=1344,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..0ecf814a28a9441e89f892eb3d63dcf8dcb0dd97 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=1344,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=14336,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=14336,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..51ad5b299eb22465fa80530d12bdd5d7a03ce398 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=14336,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=14336,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=14336,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..ee5119182556cf49434c10e56cf04e3baeb26408 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=14336,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=1792,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=1792,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..68793c77b33c4f4b97d0a4b780fcbe8043c799de --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=1792,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "5120": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "9216": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "13312": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "17408": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "25600": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "33792": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "41984": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "50176": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "58368": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..612910720ed9439e56c4af4c03f30fee224fac80 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "5120": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "9216": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "13312": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "17408": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "25600": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "33792": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "41984": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "50176": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "58368": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=2688,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=2688,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..039a10ed127b77836a7f41c03513292613852b30 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=2688,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=2688,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=2688,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..3793fcafee60bc7e8f5f12d601cb3192abfa9ca8 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=2688,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=3072,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=3072,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..51d03d8607122d7b9bc20ba48d8432d62367fa00 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=3072,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=3072,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=3072,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..26f9abd6b789e9dd0f83ec7721fd1bae8aa76bec --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=3072,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "1536": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=3200,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=3200,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..cd0cdbea0c3372674cb610870dd0b30325864549 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=3200,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,130 @@ +{ + "3328": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "768": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "1792": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2560": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "2816": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "3584": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "3840": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1280": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "2304": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=3584,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=3584,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..64be6e6591422aa0f441c3747b6c49850929652e --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=3584,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..0a6a6a73fa45e270f01ba7ebdc6d9d55bf9daad3 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "5120": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "9216": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "13312": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "17408": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "25600": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "33792": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "41984": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "50176": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "58368": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=6400,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=6400,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..ba9041d008507e31ae4179ef2bc863a49c606582 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=6400,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,130 @@ +{ + "3840": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "1792": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "3584": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "2816": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1280": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "768": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "3328": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "2560": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "2304": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..7a7508aab04599cb06641c835d8b0a14f54d0716 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..dbf9a2dd6f048d8adee290961e2aea72035f7615 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..bbb2386046b1135a2cc7ab7cb26c1d0b039bcf3a --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "1536": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=800,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=800,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..57055453aa24c831dad9ac8e37fdab707c63ef91 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=16,N=800,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,130 @@ +{ + "2048": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "1792": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "3328": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "2560": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "768": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "2816": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "2304": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2 + }, + "1280": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "3840": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "3584": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=160,N=192,device_name=NVIDIA_A800-SXM4-80GB.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=160,N=192,device_name=NVIDIA_A800-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..0611620eb33626ac891454637723eabec9aa1183 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=160,N=192,device_name=NVIDIA_A800-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=1024,device_name=AMD_Instinct_MI325X,block_shape=[128,128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=1024,device_name=AMD_Instinct_MI325X,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..43c249d2530e1fafb46f6594b73e11f33fa8ffe7 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=1024,device_name=AMD_Instinct_MI325X,block_shape=[128,128].json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=1024,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=1024,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..43c249d2530e1fafb46f6594b73e11f33fa8ffe7 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=1024,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..4dd00d110e486c0a7670504ace778fd5995de727 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..48f9697af2639bd4afd45c768c7a8e9442924e2d --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..a8c05712ba5872db5414e571fef659572e59eb80 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..f1244c61efb019183e892f848700072c3be03ef9 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..2e692a1583a4ad95f020634d55f7f4e327be5da0 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128,128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..a2ee05da1d7c632aa6b1c3ccc8af2546f987ee93 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..63e118746fd86305fd03064637ad915b72c8abcd --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..e6769604ee6bcc75c41a5bea296cb55da3c755e8 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..e6769604ee6bcc75c41a5bea296cb55da3c755e8 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..fc573cd6e8561f56830c8f8ca3033f2d2f7bc1b3 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128,128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..c6d7e96c7f0aea61a2bd57723ecb91b1c1230895 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128,128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..6fcf408755f5d3e28ff636472f89e4fa52c8da22 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_L20,dtype=fp8_w8a8,block_shape=[128,128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_L20,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..c6eabea66a39aa86f5becc04d32ee01dbe627dce --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_L20,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=512,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=512,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..e6769604ee6bcc75c41a5bea296cb55da3c755e8 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=512,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=64,device_name=NVIDIA_A800-SXM4-80GB.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=64,device_name=NVIDIA_A800-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..21f60229ff875ca6dc0cd0a3cb4fa51a845bf881 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=256,N=64,device_name=NVIDIA_A800-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=60,N=1408,device_name=AMD_Instinct_MI300X.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=60,N=1408,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..d09508b31729ddbafc8b5e1dfa88343c61e83ca9 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=60,N=1408,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "8": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=60,N=176,device_name=AMD_Instinct_MI300X.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=60,N=176,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..746463af4d56315af3c31348c2253e43c926553e --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=60,N=176,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "3072": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=60,N=352,device_name=AMD_Instinct_MI300X.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=60,N=352,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..bbdb9ad096451d6c5b9d95fae4ed02ce37715002 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=60,N=352,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "2048": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "3072": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=60,N=704,device_name=AMD_Instinct_MI300X.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=60,N=704,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..43584b1eb6b65f02fc919490b0ff4b9569de58c4 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=60,N=704,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..8cc6c643f236d2f7f9ad29354d9e469d00b20d3f --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_A800-SXM4-80GB.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_A800-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..39a9912fa4bddc48de38d5a8b4fadd71326485d2 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_A800-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..05b54639d234e20096744915c932bb2d415641a5 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..d4c9ddd12972ac0b5fd2be11a9cd1075906e3978 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..c17a4ec346915af69f10f65410450cbd66614810 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H200.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H200.json new file mode 100644 index 0000000000000000000000000000000000000000..170ae7f3fff1d09961d368853dcfc2f24c1b878c --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H200.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=2560,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=2560,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..1d9d352edebc3f228353873677d2369e6a721158 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=2560,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=2560,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=2560,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..9ad5b316750051e4a5dc39b195f6a29eeaa882f0 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=2560,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=2560,device_name=NVIDIA_H200.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=2560,device_name=NVIDIA_H200.json new file mode 100644 index 0000000000000000000000000000000000000000..2883dfd11e7f3a7db429ae0aaeb98d554f271cdf --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=2560,device_name=NVIDIA_H200.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..8abfd84a776b7be2c0920ca8d3a69c06735a08ee --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..2fc18a5e43d298ddabc57dc9f357296890fb7e29 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..be8d4a7fd23d994ac838d45bf249e39764552ccf --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H200.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H200.json new file mode 100644 index 0000000000000000000000000000000000000000..71fdd88643c6ff515808fdc2d67e9fda3ac62ae9 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H200.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..b2799ed3a866e25b78d60d92910c000ebb21ff71 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_A800-SXM4-80GB.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_A800-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..c02de2f628b716456373978f600f9fc8c4b98d24 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_A800-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_GeForce_RTX_4090,dtype=fp8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_GeForce_RTX_4090,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..3e0bc75ff87c43ad3a0b3f132e5367a8ef39619a --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_GeForce_RTX_4090,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..9f7ed6726f44e60dea5af588e441b269b6c5cc3d --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..b8d3be2313fa14025d8aeb2fd11e0d1ee997ffa6 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..21b72557e365d01011f48a046915765e67dbd13f --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H200.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H200.json new file mode 100644 index 0000000000000000000000000000000000000000..eaf32f6d76c0a6aa50a5fbc2f055a5e7c9e10b66 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H200.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..b6f1d01f886524c58932ac28c31dcf6f10788e14 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..4bf775347eccb3e76e20ea0018d71343f9848313 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..f245285bd821717f2280fa89b5f0aa6bbce2262d --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI325X.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI325X.json new file mode 100644 index 0000000000000000000000000000000000000000..3918c93b160a39e55153cbf6ff634d59e724d021 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI325X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..3f3ccdafa88f3452a695efad4cb9622d6ae79e6a --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,138 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..841044a4fc6e2da81f18b2765fa30b9de7092b3d --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=NVIDIA_H200.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=NVIDIA_H200.json new file mode 100644 index 0000000000000000000000000000000000000000..59be497fc428755f010e490277f4b945d6ed0195 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=NVIDIA_H200.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..0e5fd1eec77d7e0aca0f1e3376b2c45217ce234a --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI300X.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..d6ad63509f1573d25fbdc370b93bc39a0d00b51a --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..16e0a91baf31d9a819f27eeea719e564f399379c --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI325X.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI325X.json new file mode 100644 index 0000000000000000000000000000000000000000..d766fc062ddcce10d38bdc68ab1f700d193b8ba7 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI325X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..8323f512db0158570427a7eb0a69bd889bc2d9ed --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI300X.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..1b46cb5716514f290c6b4bc3a18dba490c799272 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..6d5b1ae5b15fc333d3cdaa5dbcacab33267339cc --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI325X.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI325X.json new file mode 100644 index 0000000000000000000000000000000000000000..ffc1b23ea90ddefd24b4b98a222e67e84442cf9c --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI325X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_A100-SXM4-40GB.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_A100-SXM4-40GB.json new file mode 100644 index 0000000000000000000000000000000000000000..f4c0f8417b384870050a95e0cf57edbdf6352b23 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_A100-SXM4-40GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..5c8185cfdeec167ec4b88de51b4b395e28769cc5 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..97c9f4445b166657ad29f1db9fc8281f9c463ec4 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..e4110a5d2e70f4d2698d4c28b1668f8e4a48688a --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_H200.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_H200.json new file mode 100644 index 0000000000000000000000000000000000000000..0883ef40582ea4f1b08c629edcd6dd9beaaf2dad --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_H200.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..81bb765d30031ca3031636d7a442dc2c2292a6a9 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI300X.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..811c77ab4109319c9f079e414bd4f7d2cf4ac9f1 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..2758e48fc4065cd05ceb6fe22574acf2b3aae633 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI325X.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI325X.json new file mode 100644 index 0000000000000000000000000000000000000000..fc31215cbae8afeee27d923e176a25daf1f2b43d --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI325X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..0bb423b28f5ab3825929a4870b96393262a9dd9f --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..55571873395464a3b58f549523905f439a8f1716 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..26bcbf26970c7a77c99e2c8eacd83eefa86967bf --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..1a0aa331933291f0c32b2cf26672771b5c7b4196 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H200.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H200.json new file mode 100644 index 0000000000000000000000000000000000000000..9952be6ba4abe434e2924e9505e834de4ce0528e --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H200.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..379ca107a9469a5e5777e4e648f3f25ffc382a0a --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..5a3f415d54142383fa7677d8192f7a4c4df66b2e --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..6cb80f48329f5718e892fd6f78000df4003832f2 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI325X.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI325X.json new file mode 100644 index 0000000000000000000000000000000000000000..de9d0aba75a7ed45fdfd92c123736f1e981a06eb --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI325X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_A100-SXM4-40GB.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_A100-SXM4-40GB.json new file mode 100644 index 0000000000000000000000000000000000000000..b41f9d443e50678334f906b44fce6d018d69500e --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_A100-SXM4-40GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..edf2a38d12ad3f420f232d2cd61ab149ad138725 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_GeForce_RTX_4090,dtype=fp8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_GeForce_RTX_4090,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..32bbadbb9eae8d36e30895d863a8d10ff65167b2 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_GeForce_RTX_4090,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..673bae2ba8ef80ed4d4930739ca7daf0e8f28ee1 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..b2100cebb7f589747430be9ca8c8db368c152d78 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..e6f753cdba35b55e8d021597dc4911bab27caa8d --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H200.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H200.json new file mode 100644 index 0000000000000000000000000000000000000000..53f3394693f060daebff9960a95ed3753e8522c9 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H200.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_L40S.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_L40S.json new file mode 100644 index 0000000000000000000000000000000000000000..d720deb4bdd73d194b1023c99e190b8fcfecdaef --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_L40S.json @@ -0,0 +1,173 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 2 + }, + "2": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 2, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 7 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 128, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 1 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 1 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 2, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 2, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 2, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 2 + }, + "192": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_ctas": 1, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 128, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 8 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_ctas": 1, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_ctas": 1, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 2 + }, + "6144": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 2 + }, + "8192": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_ctas": 1, + "num_stages": 2 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..48bb5f2ccb8e3c623593538461d8d8ba71f93a69 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI300X.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..a64d06c6d1724f736c5081a8b947fa2e6127701d --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..2c49f359c22ac8d3b8fd3b00035b70472c34ed51 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI325X.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI325X.json new file mode 100644 index 0000000000000000000000000000000000000000..c7db6c0cbd3faf615074fc8055ff744e528dd204 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI325X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..dbc624731f5cb9afcdc9213183d00d1e5edd4a00 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..cc614e635ea57327c610ce79e99ae5339614f22e --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..32c0c9da471cbe479044095e0ed14a0f54b73620 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..4dd475c02a19b380efbe8f59945a0f2a41acda11 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H200.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H200.json new file mode 100644 index 0000000000000000000000000000000000000000..2ed15f30fe603c080a2e8a32b664400a50d2153e --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H200.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..bd2c6fbc1b941cebfe8e03f62e3044f5878db0ed --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..8d7b780271854c0fbae29e92cfd3cce508ca029e --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..7a07bbf4141952c4dc6199676b1ef97d1cfabece --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI325X.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI325X.json new file mode 100644 index 0000000000000000000000000000000000000000..3a3268cc17a84d88f0a67e20e266491ad45d1777 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI325X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..f578c8d0160ac3ef85b53c8539d3675455a97173 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..918f6839620cbab1f30b0f9383a9129c2cf2cf3d --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..e341a67917d5177bacb3f6767e7b6d92539826ad --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..eb817268d41208a661d2ab9526c5108d36d44085 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H200.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H200.json new file mode 100644 index 0000000000000000000000000000000000000000..0c7062aea6c4e920cdfdc9e13fc2c0fa38aceb0e --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H200.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..cd4fb8f11b935e401103aab5117a1d1be1799c6e --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI300X.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..cf66868e9d57a541fe10fa9fff682f7a7c6dac97 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..c27ca0a36594fd12a9995fffda60a9d5003425c0 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI325X.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI325X.json new file mode 100644 index 0000000000000000000000000000000000000000..da477b1fb15eb2064cbff78b5e47ae9c54dae7a5 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI325X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..34b916e574f88c65db1dac5889d74a990dc25e9b --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..96cbc111c7fffe8b0a67c22711f9a5c8f28d8b61 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=1536,K=1536,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=1536,K=1536,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..6496a38fba8ae09b3025a75f357815b9d6a5e3f4 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=1536,K=1536,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=1536,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=1536,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..3618053b65831b95c4bb0f20ef3b9aa816b2d637 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=1536,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=1536,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=1536,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..46a982f5ee9a4bd67ce244b101c576efeeb53b78 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=1536,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=2048,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=2048,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..035ec027fa56622196b24a03a5042ce010deaebf --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=2048,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=2048,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=2048,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..8b49f2781cb54d19a2789767ebb7e8c3fb55b981 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=2048,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=2304,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=2304,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..851bc9f9f0b50b41451b929eaa518869b6a05412 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=2304,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=2304,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=2304,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..d1227c2157990216d2ca51c69ad0944017f53b6a --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=2304,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=24576,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=24576,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..77ba0d7477bdbcb036a43263e7aaa6b6913f8f4e --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=24576,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=24576,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=24576,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..1c61451fb34e52deec827f8f63c80fb15830c202 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=24576,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=24576,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=24576,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..63e661c80de6a7b1422f7a994a2ee7a4b724911c --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=24576,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=256,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=256,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..cf354037903c0d1fcd077c4647aabce026a723fb --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=256,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=3072,K=1536,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=3072,K=1536,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..0a5d7bfdba4852da9ed08d1bc27cd7d521d09965 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=3072,K=1536,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=3072,K=1536,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=3072,K=1536,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..eccb86a76df0d7302b760ab6d83a8ceb9fa9d0d9 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=3072,K=1536,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=3072,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=3072,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..cb91a279d423d0ca25197e0edd5e8c2f4da58720 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=3072,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=3072,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=3072,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..88af48431d8b8791af8df03429704606b670f1f7 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=3072,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=3072,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=3072,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..dd069726d7ed4dcbb449af243f4f4af21815f854 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=3072,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=32768,K=512,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=32768,K=512,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..7febe3d272b4bb76500f7c6b523396129fd53680 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=32768,K=512,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=32768,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=32768,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..56b939e52fac3ed53a4e0ba640c40010cb3af30a --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=32768,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=32768,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=32768,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..63d9a0bf5d79ddaaad547d44338ad4b959ad72b1 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=32768,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=36864,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=36864,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..7fa398c15a2a535401709b0f25e20f6e4b23e58e --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=36864,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=36864,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=36864,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..f15d8f64c7090bd71d0091a524c65d7818fec38e --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=36864,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=4096,K=512,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=4096,K=512,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..9d7658bfc41b2c8fd4daf3fbdf62d15936d3d546 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=4096,K=512,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=4096,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=4096,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..cd3e07804fdec10c2cfb291c1ede3ba67b753f9c --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=4096,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=4096,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=4096,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..9d5a329d7466a37c0ca68a65a089fbb99f9327a9 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=4096,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=4608,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=4608,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..03dba5ad15ba5f7f49100a5c78e8685e64334b2a --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=4608,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=4608,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=4608,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..96e1594a3eabbaedc792b84b07f05ae8752b7251 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=4608,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=4608,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=4608,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..5ffd367df833d773355590220598a3c7eceba4e0 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=4608,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=512,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=512,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..9a5ff48b8942957dde9b862aed848390dd267948 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=512,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=512,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=512,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..eabc423949a24c2a1fb2368a73e5249caf8d07df --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=512,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=576,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=576,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..386928de139ce718f28222b9c1a6555df3958491 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=576,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 5 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=576,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=576,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..51e237b91b8e775a36bcf783c078c2c1cecbcbd2 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=576,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=576,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=576,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..6280219c9ee7d26f7e2fd3625dc92d847ddc7982 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=576,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=1024,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=1024,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..40c01c0b92b4b26fe480879dda33f18c5eb59a6d --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=1024,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=1024,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=1024,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..c6fd3659799bc31e17f3577e7f0e8d7268faf1fb --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=1024,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=1152,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=1152,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..160f12ed3f95a6967439ff53bc3e3a2cdc97c700 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=1152,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=1152,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=1152,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..e5c4a1d2c94e5c7864f462e083ea5f530b8efe3f --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=1152,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=128,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=128,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..2bf5eb27e38208871d50348b170c8c74b80fc519 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=128,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "96": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=16384,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=16384,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..9c908e80406587da4d246ce4e3a8a98a14c875b1 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=16384,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=16384,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=16384,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..0a1e14cffbb2a894a701352193947d272427db0d --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=16384,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=16384,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=16384,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..15b1c93f60fc5068ba11b82b6d5924dd2024a824 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=16384,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=18432,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=18432,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..f78e7060e6840ff721d306db556636b0bbc8d9b3 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=18432,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=18432,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=18432,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..8ff12e64c172f5a5d0fbdf900728fe60b33877e2 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=18432,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=18432,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=18432,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..4532f93681e2be175b1bf94f81bfde711821cd60 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=18432,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=2048,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=2048,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..1d3ce5c94c2d9a4a1637204efb3b14f7a5579bdb --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=2048,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=2048,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=2048,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..ca7f32b9552b479dc05495792b7e426db5eb1b56 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=2048,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=2048,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=2048,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..5acea242cc0ad094cba8ee5f568ff88afb1b41ae --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=2048,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=2304,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=2304,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..3ab5796ee15b6ec8d4ab1f4ab5a594fecb30e4b4 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=2304,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=2304,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=2304,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..58cdd93e90b8c29bc7a211861711565dbeeb529a --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=2304,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=2304,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=2304,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..b72e0371d1421a1decc9d57860f83eea8f790942 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=2304,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=256,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=256,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..3cb7eaa07c745fd3aa2b3242780a7061bedac1de --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=256,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..293adce387e066fce75b6e606d4b8b6a5aa10bdb --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/configs/N=7168,K=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + } +} diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/fp8.py b/build/torch26-cxx98-cu126-aarch64-linux/moe/fp8.py new file mode 100644 index 0000000000000000000000000000000000000000..23bd7d6703104b0020671cc6ba6f78a6df37e4bf --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/fp8.py @@ -0,0 +1,63 @@ +from typing import Optional, Tuple, Union + +import torch + +from ._ops import ops + + +def is_hip() -> bool: + return torch.version.hip is not None + + +def scaled_fp8_quant( + input: torch.Tensor, + scale: Optional[torch.Tensor] = None, + num_token_padding: Optional[int] = None, + scale_ub: Optional[torch.Tensor] = None, + use_per_token_if_dynamic: bool = False, +) -> Tuple[torch.Tensor, torch.Tensor]: + """ + Quantize input tensor to FP8 and return quantized tensor and scale. + + This function supports both static and dynamic quantization: If you + provide the scale, it will use static scaling and if you omit it, + the scale will be determined dynamically. The function also allows + optional padding of the output tensors for downstream kernels that + will benefit from padding. + + Args: + input: The input tensor to be quantized to FP8 + scale: Optional scaling factor for the FP8 quantization + scale_ub: Optional upper bound for scaling factor in dynamic + per token case + num_token_padding: If specified, pad the first dimension + of the output to at least this value. + use_per_token_if_dynamic: Whether to do per_tensor or per_token + in the dynamic quantization case. + + Returns: + Tuple[torch.Tensor, torch.Tensor]: The output tensor in FP8 and + scaling factor. + """ + # This code assumes batch_dim and num_tokens are flattened + assert input.ndim == 2 + shape: Union[Tuple[int, int], torch.Size] = input.shape + # For rocm, the output fp8 dtype is torch.float_e3m3fnuz + out_dtype: torch.dtype = torch.float8_e4m3fnuz if is_hip() else torch.float8_e4m3fn + if num_token_padding: + shape = (max(num_token_padding, input.shape[0]), shape[1]) + output = torch.empty(shape, device=input.device, dtype=out_dtype) + + if scale is None: + if use_per_token_if_dynamic: + scale = torch.empty((shape[0], 1), device=input.device, dtype=torch.float32) + ops.dynamic_per_token_scaled_fp8_quant(output, input, scale, scale_ub) + else: + scale = torch.zeros(1, device=input.device, dtype=torch.float32) + ops.dynamic_scaled_fp8_quant(output, input, scale) + else: + # num_token_padding not implemented for this case + assert scale.numel() == 1 or num_token_padding is None + ops.static_scaled_fp8_quant(output, input, scale) + + return output, scale diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/fp8_utils.py b/build/torch26-cxx98-cu126-aarch64-linux/moe/fp8_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..acb4f3e3bb1a34f209fdac9ecca8c123aaf67f12 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/fp8_utils.py @@ -0,0 +1,391 @@ +import functools +import json +import logging +import os +from typing import Any, Dict, List, Optional, Tuple + +import torch +import triton +import triton.language as tl + +from .platforms import current_platform + +logger = logging.getLogger(__name__) + + +@triton.jit +def _per_token_group_quant_fp8( + # Pointers to inputs and output + y_ptr, + y_q_ptr, + y_s_ptr, + group_size, + # Avoid to divide zero + eps, + # Information for float8 + fp8_min, + fp8_max, + # Meta-parameters + BLOCK: tl.constexpr, +): + """A Triton-accelerated function to perform per-token-group + quantization on a tensor. + This function converts the tensor values into float8 values. + """ + # Map the program id to the row of X and Y it should compute. + g_id = tl.program_id(0) + y_ptr += g_id * group_size + y_q_ptr += g_id * group_size + y_s_ptr += g_id + + cols = tl.arange(0, BLOCK) # N <= BLOCK + mask = cols < group_size + + y = tl.load(y_ptr + cols, mask=mask, other=0.0).to(tl.float32) + # Quant + _absmax = tl.maximum(tl.max(tl.abs(y)), eps) + y_s = _absmax / fp8_max + y_q = tl.clamp(y / y_s, fp8_min, fp8_max).to(y_q_ptr.dtype.element_ty) + + tl.store(y_q_ptr + cols, y_q, mask=mask) + tl.store(y_s_ptr, y_s) + + +@triton.jit +def _per_token_group_quant_fp8_colmajor( + # Pointers to inputs and output + y_ptr, + y_q_ptr, + y_s_ptr, + group_size, + # Num columns of y + y_num_columns, + # Stride from one column to the next of y_s + y_s_col_stride, + # Avoid to divide zero + eps, + # Information for float8 + fp8_min, + fp8_max, + # Meta-parameters + BLOCK: tl.constexpr, +): + """A Triton-accelerated function to perform per-token-group + quantization on a tensor. + This function converts the tensor values into float8 values. + """ + # Map the program id to the row of X and Y it should compute. + g_id = tl.program_id(0) + y_ptr += g_id * group_size + y_q_ptr += g_id * group_size + + # Convert g_id the flattened block coordinate to 2D so we can index + # into the output y_scales matrix + blocks_per_row = y_num_columns // group_size + scale_col = g_id % blocks_per_row + scale_row = g_id // blocks_per_row + y_s_ptr += scale_col * y_s_col_stride + scale_row + + cols = tl.arange(0, BLOCK) # group_size <= BLOCK + mask = cols < group_size + + y = tl.load(y_ptr + cols, mask=mask, other=0.0).to(tl.float32) + # Quant + _absmax = tl.maximum(tl.max(tl.abs(y)), eps) + y_s = _absmax / fp8_max + y_q = tl.clamp(y / y_s, fp8_min, fp8_max).to(y_q_ptr.dtype.element_ty) + + tl.store(y_q_ptr + cols, y_q, mask=mask) + tl.store(y_s_ptr, y_s) + + +def per_token_group_quant_fp8( + x: torch.Tensor, + group_size: int, + eps: float = 1e-10, + dtype: Optional[torch.dtype] = None, + column_major_scales: bool = False, +) -> Tuple[torch.Tensor, torch.Tensor]: + """Function to perform per-token-group quantization on an input tensor `x`. + It converts the tensor values into signed float8 values and returns the + quantized tensor along with the scaling factor used for quantization. + Args: + x: The input tensor with ndim >= 2. + group_size: The group size used for quantization. + eps: The minimum to avoid dividing zero. + dtype: The dype of output tensor. Note that only `torch.float8_e4m3fn` + is supported for now. + Returns: + Tuple[torch.Tensor, torch.Tensor]: The quantized tensor and the + scaling factor for quantization. + """ + if dtype is None: + dtype = ( + torch.float8_e4m3fnuz if current_platform.is_rocm() else torch.float8_e4m3fn + ) + assert x.shape[-1] % group_size == 0, ( + f"the last dimension of `x` {x.shape[-1]} must be divisible " + f"by `group_size` {group_size}" + ) + assert x.is_contiguous(), "`x` must be contiguous" + + finfo = torch.finfo(dtype) + fp8_min = finfo.min + fp8_max = finfo.max + + x_q = torch.empty_like(x, device=x.device, dtype=dtype) + M = x.numel() // group_size + N = group_size + if column_major_scales: + shape = (x.shape[-1] // group_size,) + x.shape[:-1] + x_s = torch.empty(shape, device=x.device, dtype=torch.float32).permute(-1, -2) + else: + shape = x.shape[:-1] + (x.shape[-1] // group_size,) + x_s = torch.empty(shape, device=x.device, dtype=torch.float32) + + BLOCK = triton.next_power_of_2(N) + # heuristics for number of warps + num_warps = min(max(BLOCK // 256, 1), 8) + num_stages = 1 + if column_major_scales: + _per_token_group_quant_fp8_colmajor[(M,)]( + x, + x_q, + x_s, + group_size, + x.shape[1], + x_s.stride(1), + eps, + fp8_min=fp8_min, + fp8_max=fp8_max, + BLOCK=BLOCK, + num_warps=num_warps, + num_stages=num_stages, + ) + else: + _per_token_group_quant_fp8[(M,)]( + x, + x_q, + x_s, + group_size, + eps, + fp8_min=fp8_min, + fp8_max=fp8_max, + BLOCK=BLOCK, + num_warps=num_warps, + num_stages=num_stages, + ) + + return x_q, x_s + + +@triton.jit +def _w8a8_block_fp8_matmul( + # Pointers to inputs and output + A, + B, + C, + As, + Bs, + # Shape for matmul + M, + N, + K, + # Block size for block-wise quantization + group_n, + group_k, + # Stride for inputs and output + stride_am, + stride_ak, + stride_bk, + stride_bn, + stride_cm, + stride_cn, + stride_As_m, + stride_As_k, + stride_Bs_k, + stride_Bs_n, + # Meta-parameters + BLOCK_SIZE_M: tl.constexpr, + BLOCK_SIZE_N: tl.constexpr, + BLOCK_SIZE_K: tl.constexpr, + GROUP_SIZE_M: tl.constexpr, +): + """Triton-accelerated function used to perform linear operations (dot + product) on input tensors `A` and `B` with block-wise quantization, and + store the result in output tensor `C`. + """ + + pid = tl.program_id(axis=0) + num_pid_m = tl.cdiv(M, BLOCK_SIZE_M) + num_pid_n = tl.cdiv(N, BLOCK_SIZE_N) + num_pid_in_group = GROUP_SIZE_M * num_pid_n + group_id = pid // num_pid_in_group + first_pid_m = group_id * GROUP_SIZE_M + group_size_m = min(num_pid_m - first_pid_m, GROUP_SIZE_M) + pid_m = first_pid_m + (pid % group_size_m) + pid_n = (pid % num_pid_in_group) // group_size_m + + offs_am = (pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)) % M + offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)) % N + offs_k = tl.arange(0, BLOCK_SIZE_K) + a_ptrs = A + (offs_am[:, None] * stride_am + offs_k[None, :] * stride_ak) + b_ptrs = B + (offs_k[:, None] * stride_bk + offs_bn[None, :] * stride_bn) + + As_ptrs = As + offs_am * stride_As_m + offs_bsn = offs_bn // group_n + Bs_ptrs = Bs + offs_bsn * stride_Bs_n + + accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32) + for k in range(0, tl.cdiv(K, BLOCK_SIZE_K)): + a = tl.load(a_ptrs, mask=offs_k[None, :] < K - k * BLOCK_SIZE_K, other=0.0) + b = tl.load(b_ptrs, mask=offs_k[:, None] < K - k * BLOCK_SIZE_K, other=0.0) + + k_start = k * BLOCK_SIZE_K + offs_ks = k_start // group_k + a_s = tl.load(As_ptrs + offs_ks * stride_As_k) + b_s = tl.load(Bs_ptrs + offs_ks * stride_Bs_k) + + accumulator += tl.dot(a, b) * a_s[:, None] * b_s[None, :] + a_ptrs += BLOCK_SIZE_K * stride_ak + b_ptrs += BLOCK_SIZE_K * stride_bk + + if C.dtype.element_ty == tl.bfloat16: + c = accumulator.to(tl.bfloat16) + elif C.dtype.element_ty == tl.float16: + c = accumulator.to(tl.float16) + else: + c = accumulator.to(tl.float32) + + offs_cm = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M) + offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N) + c_ptrs = C + stride_cm * offs_cm[:, None] + stride_cn * offs_cn[None, :] + c_mask = (offs_cm[:, None] < M) & (offs_cn[None, :] < N) + tl.store(c_ptrs, c, mask=c_mask) + + +@functools.lru_cache +def get_w8a8_block_fp8_configs( + N: int, K: int, block_n: int, block_k: int +) -> Optional[Dict[int, Any]]: + """ + Return optimized configurations for the w8a8 block fp8 kernel. + The return value will be a dictionary that maps an irregular grid of + batch sizes to configurations of the w8a8 block fp8 kernel. To evaluate the + kernel on a given batch size bs, the closest batch size in the grid should + be picked and the associated configuration chosen to invoke the kernel. + """ + + # First look up if an optimized configuration is available in the configs + # directory + device_name = current_platform.get_device_name().replace(" ", "_") + json_file_name = f"N={N},K={K},device_name={device_name},dtype=fp8_w8a8,block_shape=[{block_n}, {block_k}].json" # noqa: E501 + + config_file_path = os.path.join( + os.path.dirname(os.path.realpath(__file__)), "configs", json_file_name + ) + if os.path.exists(config_file_path): + with open(config_file_path) as f: + logger.info( + "Using configuration from %s for W8A8 Block FP8 kernel.", + config_file_path, + ) + # If a configuration has been found, return it + return {int(key): val for key, val in json.load(f).items()} + + # If no optimized configuration is available, we will use the default + # configuration + logger.warning( + "Using default W8A8 Block FP8 kernel config. Performance might " + "be sub-optimal! Config file not found at %s", + config_file_path, + ) + return None + + +def w8a8_block_fp8_matmul( + A: torch.Tensor, + B: torch.Tensor, + As: torch.Tensor, + Bs: torch.Tensor, + block_size: List[int], + output_dtype: torch.dtype = torch.float16, +) -> torch.Tensor: + """This function performs matrix multiplication with block-wise + quantization. + It takes two input tensors `A` and `B` with scales `As` and `Bs`. + The output is returned in the specified `output_dtype`. + Args: + A: The input tensor, e.g., activation. + B: The input tensor, e.g., weight. + As: The per-token-group quantization scale for `A`. + Bs: The per-block quantization scale for `B`. + block_size: The block size for per-block quantization. It should + be 2-dim, e.g., [128, 128]. + output_dytpe: The dtype of the returned tensor. + Returns: + torch.Tensor: The result of matmul. + """ + assert len(block_size) == 2 + block_n, block_k = block_size[0], block_size[1] + + assert A.shape[-1] == B.shape[-1] + assert A.shape[:-1] == As.shape[:-1] and A.is_contiguous() + assert triton.cdiv(A.shape[-1], block_k) == As.shape[-1] + M = A.numel() // A.shape[-1] + + assert B.ndim == 2 and B.is_contiguous() and Bs.ndim == 2 + N, K = B.shape + assert triton.cdiv(N, block_n) == Bs.shape[0] + assert triton.cdiv(K, block_k) == Bs.shape[1] + + C_shape = A.shape[:-1] + (N,) + C = A.new_empty(C_shape, dtype=output_dtype) + + configs = get_w8a8_block_fp8_configs(N, K, block_size[0], block_size[1]) + if configs: + # Get the optimal config if there is one + config = configs[min(configs.keys(), key=lambda x: abs(x - M))] + else: + # Default config + # Block-wise quant: BLOCK_SIZE_N must be divisible by block_size[0] + # BLOCK_SIZE_K must be divisible by block_size[1] + config = { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": block_size[0], + "BLOCK_SIZE_K": block_size[1], + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2, + } + + def grid(META): + return ( + triton.cdiv(M, META["BLOCK_SIZE_M"]) * triton.cdiv(N, META["BLOCK_SIZE_N"]), + ) + + _w8a8_block_fp8_matmul[grid]( + A, + B, + C, + As, + Bs, + M, + N, + K, + block_n, + block_k, + A.stride(-2), + A.stride(-1), + B.stride(1), + B.stride(0), + C.stride(-2), + C.stride(-1), + As.stride(-2), + As.stride(-1), + Bs.stride(1), + Bs.stride(0), + **config, + ) + + return C diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/fused_marlin_moe.py b/build/torch26-cxx98-cu126-aarch64-linux/moe/fused_marlin_moe.py new file mode 100644 index 0000000000000000000000000000000000000000..b3e0a5c24599730faf973fad3cf3fb6031a30522 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/fused_marlin_moe.py @@ -0,0 +1,413 @@ +"""Fused MoE utilities for GPTQ.""" + +import functools +from typing import TYPE_CHECKING, Optional + +import torch + +from ._ops import add_op_namespace_prefix, ops +from .fused_moe import fused_topk, moe_align_block_size, try_get_optimal_moe_config +from .scalar_type import ScalarType, scalar_types + +# neuron has torch version that doesn't even have impl_abstract +if TYPE_CHECKING: + + def register_fake(fn): + return lambda name: fn + +else: + try: + from torch.library import register_fake + except ImportError: + from torch.library import impl_abstract as register_fake + + +def get_scalar_type(num_bits: int, has_zp: bool): + if has_zp: + assert num_bits == 4 + return scalar_types.uint4 + else: + return scalar_types.uint4b8 if num_bits == 4 else scalar_types.uint8b128 + + +def single_marlin_moe( + hidden_states: torch.Tensor, + w: torch.Tensor, + scales: torch.Tensor, + gating_output: torch.Tensor, + topk: int, + renormalize: bool, + g_idx: Optional[torch.Tensor] = None, + sort_indices: Optional[torch.Tensor] = None, + w_zeros: Optional[torch.Tensor] = None, + num_bits: int = 8, + is_k_full: bool = True, +) -> torch.Tensor: + """ + This function computes the multiplication of hidden_states with expert + weights used in Marlin MoE, using weights w and top-k gating mechanism. + Its purpose is testing and debugging the fused MoE kernel. + + Parameters: + - hidden_states (torch.Tensor): The input tensor to the Marlin Mul. + - w (torch.Tensor): The set of expert weights. + - scales (torch.Tensor): The quantization scales. + - gating_output (torch.Tensor): The output of the gating operation + (before softmax). + - g_idx (Optional[torch.Tensor]): Optional act_order indices. + - sort_indices (Optional[torch.Tensor]): Optional act_order input + permutation. + - topk (int): The number of top-k experts to select. + - renormalize (bool): If True, renormalize the top-k weights to sum to 1. + - w_zeros (Optional[torch.Tensor]): Optional zero points to be used for w. + - num_bits (bool): The number of bits in expert weights quantization. + + Returns: + - torch.Tensor: The output tensor after applying the MoE layer. + """ + # Check constraints. + assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch" + assert hidden_states.shape[1] == w.shape[1] * 16, "Hidden size mismatch" + assert gating_output.shape[1] == w.shape[0], "Number of experts mismatch" + assert hidden_states.is_contiguous(), "Hidden_states must be contiguous" + assert w.is_contiguous(), "Expert weights must be contiguous" + assert hidden_states.dtype == torch.float16 + assert num_bits in [4, 8] + + M, K = hidden_states.shape + E = w.shape[0] + N = w.shape[2] // (num_bits // 2) + + topk_weights, topk_ids = fused_topk(hidden_states, gating_output, topk, renormalize) + + # This might not be an optimal config for a single MMM + get_config_func = functools.partial( + try_get_optimal_moe_config, + w.shape, + w.shape, + topk_ids.shape[1], + None, + is_marlin=True, + ) + config = get_config_func(M) + + block_size_m = config["BLOCK_SIZE_M"] + + sorted_token_ids, _, _ = moe_align_block_size(topk_ids, block_size_m, E) + + max_workspace_size = (N // 64) * 16 + workspace = torch.zeros( + max_workspace_size, + dtype=torch.int, + device=hidden_states.device, + requires_grad=False, + ) + + has_zero_point = w_zeros is not None + if w_zeros is None: + w_zeros = torch.empty( + (0, 0), + dtype=hidden_states.dtype, + device=hidden_states.device, + requires_grad=False, + ) + + if g_idx is None: + g_idx = torch.empty( + (0, 0), dtype=torch.int32, device=hidden_states.device, requires_grad=False + ) + + if sort_indices is None: + sort_indices = torch.empty( + (0), dtype=torch.int32, device=hidden_states.device, requires_grad=False + ) + + scalar_type = get_scalar_type(num_bits, has_zero_point) + + intermediate_cache = ops.marlin_gemm_moe( + hidden_states, + w, + sorted_token_ids, + topk_weights, + topk_ids, + scales, + w_zeros, + g_idx, + sort_indices, + workspace, + scalar_type.id, + M, + N, + K, + is_k_full, + E, + topk, + block_size_m, + True, + False, + ) + + return torch.sum(intermediate_cache.view(*intermediate_cache.shape), dim=1) + + +if hasattr(ops, "single_marlin_gemm_moe"): + + @register_fake(add_op_namespace_prefix("single_marlin_gemm_moe")) + def single_marlin_moe_fake( + hidden_states: torch.Tensor, + w: torch.Tensor, + scales: torch.Tensor, + gating_output: torch.Tensor, + topk: int, + renormalize: bool, + g_idx: Optional[torch.Tensor] = None, + sort_indices: Optional[torch.Tensor] = None, + w_zeros: Optional[torch.Tensor] = None, + num_bits: int = 8, + is_k_full: bool = True, + ) -> torch.Tensor: + return torch.empty_like(hidden_states) + + +def fused_marlin_moe( + hidden_states: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + w1_scale: torch.Tensor, + w2_scale: torch.Tensor, + gating_output: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + g_idx1: Optional[torch.Tensor] = None, + g_idx2: Optional[torch.Tensor] = None, + sort_indices1: Optional[torch.Tensor] = None, + sort_indices2: Optional[torch.Tensor] = None, + w1_zeros: Optional[torch.Tensor] = None, + w2_zeros: Optional[torch.Tensor] = None, + num_bits: int = 8, + is_k_full: bool = True, +) -> torch.Tensor: + """ + This function computes a Mixture of Experts (MoE) layer using two sets of + weights, w1 and w2, and top-k gating mechanism. + + Parameters: + - hidden_states (torch.Tensor): The input tensor to the MoE layer. + - w1 (torch.Tensor): The first set of expert weights. + - w2 (torch.Tensor): The second set of expert weights. + - w1_scale (torch.Tensor): Scale to be used for w1. + - w2_scale (torch.Tensor): Scale to be used for w2. + - gating_output (torch.Tensor): The output of the gating operation + (before softmax). + - g_idx1 (Optional[torch.Tensor]): The first set of act_order indices. + - g_idx2 (Optional[torch.Tensor]): The second set of act_order indices. + - sort_indices1 (Optional[torch.Tensor]): The first act_order input + permutation. + - sort_indices2 (Optional[torch.Tensor]): The second act_order input + permutation. + - topk_weights (torch.Tensor): Top-k weights. + - topk_ids (torch.Tensor): Indices of topk-k elements. + - w1_zeros (Optional[torch.Tensor]): Optional zero points to be used for w1. + - w2_zeros (Optional[torch.Tensor]): Optional zero points to be used for w2. + - num_bits (bool): The number of bits in expert weights quantization. + + Returns: + - torch.Tensor: The output tensor after applying the MoE layer. + """ + # Check constraints. + assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch" + assert hidden_states.shape[1] == w1.shape[1] * 16, "Hidden size mismatch w1" + assert hidden_states.shape[1] == w2.shape[2] // ( + num_bits // 2 + ), "Hidden size mismatch w2" + assert gating_output.shape[1] == w1.shape[0], "Number of experts mismatch" + assert hidden_states.is_contiguous(), "Hidden_states must be contiguous" + assert w1.is_contiguous(), "Expert weights1 must be contiguous" + assert w2.is_contiguous(), "Expert weights2 must be contiguous" + assert hidden_states.dtype == torch.float16 + assert num_bits in [4, 8] + + has_no_act_order = ( + g_idx1 is None + and g_idx2 is None + and sort_indices1 is None + and sort_indices2 is None + ) + has_all_act_order = ( + g_idx1 is not None + and g_idx2 is not None + and sort_indices1 is not None + and sort_indices2 is not None + ) + assert has_no_act_order or has_all_act_order, ( + "g_idx and sorted_indices " "must be all not None or must be all None" + ) + + has_no_zp = w1_zeros is None and w2_zeros is None + has_all_zp = w1_zeros is not None and w2_zeros is not None + assert has_no_zp or has_all_zp, ( + "zero points must be both not None or " "must be both None" + ) + + M, K = hidden_states.shape + E = w1.shape[0] + N = w2.shape[1] * 16 + topk = topk_ids.shape[1] + + get_config_func = functools.partial( + try_get_optimal_moe_config, + w1.shape, + w2.shape, + topk_ids.shape[1], + None, + is_marlin=True, + ) + config = get_config_func(M) + + block_size_m = config["BLOCK_SIZE_M"] + + sorted_token_ids, _, _ = moe_align_block_size(topk_ids, block_size_m, E) + + max_workspace_size = (max(2 * N, K) // 64) * 16 + workspace = torch.zeros( + max_workspace_size, dtype=torch.int, device="cuda", requires_grad=False + ) + + if has_no_zp: + w1_zeros = torch.empty( + (0, 0), + dtype=hidden_states.dtype, + device=hidden_states.device, + requires_grad=False, + ) + w2_zeros = torch.empty( + (0, 0), + dtype=hidden_states.dtype, + device=hidden_states.device, + requires_grad=False, + ) + + if has_no_act_order: + g_idx1 = torch.empty( + (0, 0), dtype=torch.int32, device=hidden_states.device, requires_grad=False + ) + g_idx2 = torch.empty( + (0, 0), dtype=torch.int32, device=hidden_states.device, requires_grad=False + ) + sort_indices1 = torch.empty( + (0), dtype=torch.int32, device=hidden_states.device, requires_grad=False + ) + sort_indices2 = torch.empty( + (0, 0), dtype=torch.int32, device=hidden_states.device, requires_grad=False + ) + + scalar_type1 = get_scalar_type(num_bits, has_all_zp) + scalar_type2 = get_scalar_type(num_bits, has_all_zp) + + intermediate_cache2 = torch.empty( + (M * topk_ids.shape[1], N), + device=hidden_states.device, + dtype=hidden_states.dtype, + ) + + intermediate_cache1 = ops.marlin_gemm_moe( + hidden_states, + w1, + sorted_token_ids, + topk_weights, + topk_ids, + w1_scale, + w1_zeros, + g_idx1, + sort_indices1, + workspace, + scalar_type1.id, + M, + 2 * N, + K, + is_k_full, + E, + topk, + block_size_m, + True, + False, + ) + + ops.silu_and_mul(intermediate_cache2, intermediate_cache1.view(-1, 2 * N)) + + intermediate_cache3 = ops.marlin_gemm_moe( + intermediate_cache2, + w2, + sorted_token_ids, + topk_weights, + topk_ids, + w2_scale, + w2_zeros, + g_idx2, + sort_indices2, + workspace, + scalar_type2.id, + M, + K, + N, + is_k_full, + E, + topk, + block_size_m, + False, + True, + ) + + return torch.sum(intermediate_cache3.view(*intermediate_cache3.shape), dim=1) + + +if hasattr(ops, "fused_marlin_moe"): + + @register_fake(add_op_namespace_prefix("fused_marlin_moe")) + def fused_marlin_moe_fake( + hidden_states: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + w1_scale: torch.Tensor, + w2_scale: torch.Tensor, + gating_output: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + g_idx1: Optional[torch.Tensor] = None, + g_idx2: Optional[torch.Tensor] = None, + sort_indices1: Optional[torch.Tensor] = None, + sort_indices2: Optional[torch.Tensor] = None, + w1_zeros: Optional[torch.Tensor] = None, + w2_zeros: Optional[torch.Tensor] = None, + num_bits: int = 8, + is_k_full: bool = True, + ) -> torch.Tensor: + return torch.empty_like(hidden_states) + + +if hasattr(ops, "marlin_gemm_moe"): + + @register_fake(add_op_namespace_prefix("marlin_gemm_moe")) + def marlin_gemm_moe_fake( + a: torch.Tensor, + b_q_weights: torch.Tensor, + sorted_ids: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + b_scales: torch.Tensor, + b_zero_points: torch.Tensor, + g_idx: torch.Tensor, + perm: torch.Tensor, + workspace: torch.Tensor, + b_q_type: ScalarType, + size_m: torch.SymInt, + size_n: torch.SymInt, + size_k: torch.SymInt, + is_k_full: bool, + num_experts: int, + topk: int, + moe_block_size: int, + replicate_input: bool, + apply_weights: bool, + ) -> torch.Tensor: + return torch.empty((size_m, topk, size_n), dtype=a.dtype, device=a.device) diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/fused_moe.py b/build/torch26-cxx98-cu126-aarch64-linux/moe/fused_moe.py new file mode 100644 index 0000000000000000000000000000000000000000..f8f0586c2a58d1fb68f7d2eaed47a48f165d1e8d --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/fused_moe.py @@ -0,0 +1,1574 @@ +# SPDX-License-Identifier: Apache-2.0 +"""Fused MoE kernel.""" + +import functools +import json +import logging +import os +from typing import Any, Callable, Dict, List, Optional, Tuple + +import torch +import triton +import triton.language as tl + +from ._ops import ops +from .fp8 import scaled_fp8_quant +from .fp8_utils import per_token_group_quant_fp8 +from .moe_align_block_size import moe_align_block_size +from .platforms import current_platform +from .utils.custom_ops import direct_register_custom_op + +logger = logging.getLogger(__name__) + + +VLLM_FUSED_MOE_CHUNK_SIZE = int(os.getenv("VLLM_FUSED_MOE_CHUNK_SIZE", "32768")) +VLLM_ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON: bool = bool( + int(os.getenv("VLLM_ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON", "0")) +) + + +def cdiv(a: int, b: int) -> int: + """Ceiling division.""" + return -(a // -b) + + +def _fp8_quantize( + A: torch.Tensor, + A_scale: Optional[torch.Tensor], + block_shape: Optional[List[int]], +) -> Tuple[torch.Tensor, torch.Tensor]: + """ + Perform fp8 quantization on the inputs. If a block_shape + is provided, the output will be blocked. + """ + if block_shape is None: + A, A_scale = scaled_fp8_quant(A, A_scale) + else: + assert len(block_shape) == 2 + _, block_k = block_shape[0], block_shape[1] + A, A_scale = per_token_group_quant_fp8(A, block_k) + assert cdiv(A.shape[-1], block_k) == A_scale.shape[-1] + return A, A_scale + + +@triton.jit +def write_zeros_to_output( + c_ptr, + stride_cm, + stride_cn, + pid_n, + N, + offs_token, + token_mask, + BLOCK_SIZE_M, + BLOCK_SIZE_N, + compute_type, +): + accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=compute_type) + offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N) + c_ptrs = c_ptr + stride_cm * offs_token[:, None] + stride_cn * offs_cn[None, :] + c_mask = token_mask[:, None] & (offs_cn[None, :] < N) + tl.store(c_ptrs, accumulator, mask=c_mask) + + +@triton.jit +def fused_moe_kernel_gptq_awq( + # Pointers to matrices + a_ptr, + b_ptr, + c_ptr, + b_scale_ptr, + b_zp_ptr, + topk_weights_ptr, + sorted_token_ids_ptr, + expert_ids_ptr, + num_tokens_post_padded_ptr, + # Matrix dimensions + N: tl.constexpr, + K: tl.constexpr, + EM, + num_valid_tokens, + # The stride variables represent how much to increase the ptr by when + # moving by 1 element in a particular dimension. E.g. `stride_am` is + # how much to increase `a_ptr` by to get the element one row down + # (A has M rows). + stride_am, + stride_ak, + stride_be, + stride_bk, + stride_bn, + stride_cm, + stride_cn, + stride_bse, + stride_bsk, + stride_bsn, + stride_bze, + stride_bzk, + stride_bzn, + block_k_diviable: tl.constexpr, + group_size: tl.constexpr, + # Meta-parameters + BLOCK_SIZE_M: tl.constexpr, + BLOCK_SIZE_N: tl.constexpr, + BLOCK_SIZE_K: tl.constexpr, + GROUP_SIZE_M: tl.constexpr, + MUL_ROUTED_WEIGHT: tl.constexpr, + top_k: tl.constexpr, + compute_type: tl.constexpr, + has_zp: tl.constexpr, + use_int4_w4a16: tl.constexpr, + use_int8_w8a16: tl.constexpr, +): + """ + Implements the fused computation for a Mixture of Experts (MOE) using + token and expert matrices. + + Key Parameters: + - A: The input tensor representing tokens with shape (*, K), where '*' can + be any shape representing batches and K is the feature dimension of + each token. + - B: The stacked MOE weight tensor with shape (E, N, K), where E is + the number of experts, K is the input feature dimension, and N is + the output feature dimension. + - C: The output cache tensor with shape (M, topk, N), where M is the + total number of tokens post padding, topk is the number of times + each token is repeated, and N is the output feature dimension. + - sorted_token_ids: A tensor containing the sorted indices of tokens, + repeated topk times and arranged by the expert index they are + assigned to. + - expert_ids: A tensor containing the indices of the expert for each + block. It determines which expert matrix from B should be used for + each block in A. + This kernel performs the multiplication of a token by its corresponding + expert matrix as determined by `expert_ids`. The sorting of + `sorted_token_ids` by expert index and padding ensures divisibility by + BLOCK_SIZE_M, which is necessary to maintain consistency in block matrix + multiplication across different blocks processed by the same expert. + """ + # ----------------------------------------------------------- + # Map program ids `pid` to the block of C it should compute. + # This is done in a grouped ordering to promote L2 data reuse. + pid = tl.program_id(axis=0) + num_pid_m = tl.cdiv(EM, BLOCK_SIZE_M) + num_pid_n = tl.cdiv(N, BLOCK_SIZE_N) + num_pid_in_group = GROUP_SIZE_M * num_pid_n + group_id = pid // num_pid_in_group + first_pid_m = group_id * GROUP_SIZE_M + group_size_m = min(num_pid_m - first_pid_m, GROUP_SIZE_M) + pid_m = first_pid_m + ((pid % num_pid_in_group) % group_size_m) + pid_n = (pid % num_pid_in_group) // group_size_m + + # ---------------------------------------------------------- + # Create pointers for the first blocks of A and B. + # We will advance this pointer as we move in the K direction + # and accumulate + # `a_ptrs` is a block of [BLOCK_SIZE_M, BLOCK_SIZE_K] pointers + # `b_ptrs` is a block of [BLOCK_SIZE_K, BLOCK_SIZE_N] pointers + num_tokens_post_padded = tl.load(num_tokens_post_padded_ptr) + if pid_m * BLOCK_SIZE_M >= num_tokens_post_padded: + return + offs_token_id = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M).to(tl.int64) + offs_token = tl.load(sorted_token_ids_ptr + offs_token_id) + token_mask = offs_token < num_valid_tokens + + off_experts = tl.load(expert_ids_ptr + pid_m).to(tl.int64) + if off_experts == -1: + # ----------------------------------------------------------- + # Write back zeros to the output when the expert is not + # in the current expert parallel rank. + write_zeros_to_output( + c_ptr, + stride_cm, + stride_cn, + pid_n, + N, + offs_token, + token_mask, + BLOCK_SIZE_M, + BLOCK_SIZE_N, + compute_type, + ) + return + + offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N).to(tl.int64)) % N + offs_k = tl.arange(0, BLOCK_SIZE_K) + a_ptrs = a_ptr + ( + offs_token[:, None] // top_k * stride_am + offs_k[None, :] * stride_ak + ) + + if use_int4_w4a16: + b_ptrs = ( + b_ptr + + off_experts * stride_be + + (offs_k[:, None] // 2) * stride_bk + + offs_bn[None, :] * stride_bn + ) + b_shifter = (offs_k[:, None] % 2) * 4 + elif use_int8_w8a16: + b_ptrs = ( + b_ptr + + off_experts * stride_be + + offs_k[:, None] * stride_bk + + offs_bn[None, :] * stride_bn + ) + + if not has_zp and use_int4_w4a16: + b_zp_num = 8 + if not has_zp and use_int8_w8a16: + b_zp_num = 128 + elif has_zp and use_int4_w4a16: + b_zp_shifter = (offs_bn[None, :] % 2) * 4 + + # ----------------------------------------------------------- + # Iterate to compute a block of the C matrix. + # We accumulate into a `[BLOCK_SIZE_M, BLOCK_SIZE_N]` block + # of fp32 values for higher accuracy. + # `accumulator` will be converted back to fp16 after the loop. + accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32) + for k in range(0, tl.cdiv(K, BLOCK_SIZE_K)): + # Load the next block of A and B, generate a mask by checking the + # K dimension. + + if not block_k_diviable: + k_mask = offs_k[:, None] < K - k * BLOCK_SIZE_K + k_other = 0.0 + else: + k_mask = None + k_other = None + + a = tl.load( + a_ptrs, + mask=token_mask[:, None] & (offs_k[None, :] < K - k * BLOCK_SIZE_K), + other=0.0, + ) + b = tl.load(b_ptrs) + if use_int4_w4a16: + b = (b >> b_shifter) & 0xF + + b_scale_ptrs = ( + b_scale_ptr + + off_experts * stride_bse + + offs_bn[None, :] * stride_bsn + + ((offs_k[:, None] + BLOCK_SIZE_K * k) // group_size) * stride_bsk + ) + b_scale = tl.load(b_scale_ptrs, mask=k_mask, other=k_other) + b_scale = b_scale.to(tl.float32) + + if has_zp and use_int4_w4a16: + offs_k_true = (offs_k[:, None] + BLOCK_SIZE_K * k) // group_size + b_zp_ptrs = ( + b_zp_ptr + + off_experts * stride_bze + + (offs_bn[None, :] // 2) * stride_bzn + + offs_k_true * stride_bzk + ) + b_zp = tl.load(b_zp_ptrs, mask=k_mask, other=k_other) + b_zp = (b_zp >> b_zp_shifter) & 0xF + b_zp = b_zp.to(tl.float32) + elif has_zp and use_int8_w8a16: + offs_k_true = (offs_k[:, None] + BLOCK_SIZE_K * k) // group_size + b_zp_ptrs = ( + b_zp_ptr + + off_experts * stride_bze + + offs_bn[None, :] * stride_bzn + + offs_k_true * stride_bzk + ) + b_zp = tl.load(b_zp_ptrs, mask=k_mask, other=k_other) + b_zp = b_zp.to(tl.float32) + + # We accumulate along the K dimension. + if has_zp: + b = ((b.to(tl.float32) - b_zp) * b_scale).to(compute_type) + else: + b = ((b.to(tl.float32) - b_zp_num) * b_scale).to(compute_type) + accumulator = tl.dot(a, b, acc=accumulator) + + # Advance the ptrs to the next K block. + a_ptrs += BLOCK_SIZE_K * stride_ak + if use_int4_w4a16: + b_ptrs += (BLOCK_SIZE_K // 2) * stride_bk + else: + b_ptrs += BLOCK_SIZE_K * stride_bk + + if MUL_ROUTED_WEIGHT: + moe_weight = tl.load(topk_weights_ptr + offs_token, mask=token_mask, other=0) + accumulator = accumulator * moe_weight[:, None] + + accumulator = accumulator.to(compute_type) + # ----------------------------------------------------------- + # Write back the block of the output + offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N) + c_ptrs = c_ptr + stride_cm * offs_token[:, None] + stride_cn * offs_cn[None, :] + c_mask = token_mask[:, None] & (offs_cn[None, :] < N) + tl.store(c_ptrs, accumulator, mask=c_mask) + + +@triton.jit +def fused_moe_kernel( + # Pointers to matrices + a_ptr, + b_ptr, + c_ptr, + a_scale_ptr, + b_scale_ptr, + topk_weights_ptr, + sorted_token_ids_ptr, + expert_ids_ptr, + num_tokens_post_padded_ptr, + # Matrix dimensions + N, + K, + EM, + num_valid_tokens, + # The stride variables represent how much to increase the ptr by when + # moving by 1 element in a particular dimension. E.g. `stride_am` is + # how much to increase `a_ptr` by to get the element one row down + # (A has M rows). + stride_am, + stride_ak, + stride_be, + stride_bk, + stride_bn, + stride_cm, + stride_cn, + stride_asm, + stride_ask, + stride_bse, + stride_bsk, + stride_bsn, + # Block size for block-wise quantization + group_n: tl.constexpr, + group_k: tl.constexpr, + # Meta-parameters + BLOCK_SIZE_M: tl.constexpr, + BLOCK_SIZE_N: tl.constexpr, + BLOCK_SIZE_K: tl.constexpr, + GROUP_SIZE_M: tl.constexpr, + MUL_ROUTED_WEIGHT: tl.constexpr, + top_k: tl.constexpr, + compute_type: tl.constexpr, + use_fp8_w8a8: tl.constexpr, + use_int8_w8a16: tl.constexpr, +): + """ + Implements the fused computation for a Mixture of Experts (MOE) using + token and expert matrices. + + Key Parameters: + - A: The input tensor representing tokens with shape (*, K), where '*' can + be any shape representing batches and K is the feature dimension of + each token. + - B: The stacked MOE weight tensor with shape (E, N, K), where E is + the number of experts, K is the input feature dimension, and N is + the output feature dimension. + - C: The output cache tensor with shape (M, topk, N), where M is the + total number of tokens post padding, topk is the number of times + each token is repeated, and N is the output feature dimension. + - sorted_token_ids: A tensor containing the sorted indices of tokens, + repeated topk times and arranged by the expert index they are + assigned to. + - expert_ids: A tensor containing the indices of the expert for each + block. It determines which expert matrix from B should be used for + each block in A. + This kernel performs the multiplication of a token by its corresponding + expert matrix as determined by `expert_ids`. The sorting of + `sorted_token_ids` by expert index and padding ensures divisibility by + BLOCK_SIZE_M, which is necessary to maintain consistency in block matrix + multiplication across different blocks processed by the same expert. + """ + # ----------------------------------------------------------- + # Map program ids `pid` to the block of C it should compute. + # This is done in a grouped ordering to promote L2 data reuse. + pid = tl.program_id(axis=0) + num_pid_m = tl.cdiv(EM, BLOCK_SIZE_M) + num_pid_n = tl.cdiv(N, BLOCK_SIZE_N) + num_pid_in_group = GROUP_SIZE_M * num_pid_n + group_id = pid // num_pid_in_group + first_pid_m = group_id * GROUP_SIZE_M + group_size_m = min(num_pid_m - first_pid_m, GROUP_SIZE_M) + pid_m = first_pid_m + ((pid % num_pid_in_group) % group_size_m) + pid_n = (pid % num_pid_in_group) // group_size_m + + # ---------------------------------------------------------- + # Create pointers for the first blocks of A and B. + # We will advance this pointer as we move in the K direction + # and accumulate + # `a_ptrs` is a block of [BLOCK_SIZE_M, BLOCK_SIZE_K] pointers + # `b_ptrs` is a block of [BLOCK_SIZE_K, BLOCK_SIZE_N] pointers + num_tokens_post_padded = tl.load(num_tokens_post_padded_ptr) + if pid_m * BLOCK_SIZE_M >= num_tokens_post_padded: + return + offs_token_id = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M).to(tl.int64) + offs_token = tl.load(sorted_token_ids_ptr + offs_token_id) + token_mask = offs_token < num_valid_tokens + + off_experts = tl.load(expert_ids_ptr + pid_m).to(tl.int64) + if off_experts == -1: + # ----------------------------------------------------------- + # Write back zeros to the output when the expert is not + # in the current expert parallel rank. + write_zeros_to_output( + c_ptr, + stride_cm, + stride_cn, + pid_n, + N, + offs_token, + token_mask, + BLOCK_SIZE_M, + BLOCK_SIZE_N, + compute_type, + ) + return + + offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N).to(tl.int64)) % N + offs_k = tl.arange(0, BLOCK_SIZE_K) + a_ptrs = a_ptr + ( + offs_token[:, None] // top_k * stride_am + offs_k[None, :] * stride_ak + ) + + b_ptrs = ( + b_ptr + + off_experts * stride_be + + (offs_k[:, None] * stride_bk + offs_bn[None, :] * stride_bn) + ) + if use_int8_w8a16: + b_scale_ptrs = ( + b_scale_ptr + off_experts * stride_bse + offs_bn[None, :] * stride_bsn + ) + b_scale = tl.load(b_scale_ptrs) + + if use_fp8_w8a8: + if group_k > 0 and group_n > 0: + a_scale_ptrs = a_scale_ptr + (offs_token // top_k) * stride_asm + offs_bsn = offs_bn // group_n + b_scale_ptrs = ( + b_scale_ptr + off_experts * stride_bse + offs_bsn * stride_bsn + ) + else: + a_scale = tl.load(a_scale_ptr) + b_scale = tl.load(b_scale_ptr + off_experts) + + # ----------------------------------------------------------- + # Iterate to compute a block of the C matrix. + # We accumulate into a `[BLOCK_SIZE_M, BLOCK_SIZE_N]` block + # of fp32 values for higher accuracy. + # `accumulator` will be converted back to fp16 after the loop. + accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32) + for k in range(0, tl.cdiv(K, BLOCK_SIZE_K)): + # Load the next block of A and B, generate a mask by checking the + # K dimension. + a = tl.load( + a_ptrs, + mask=token_mask[:, None] & (offs_k[None, :] < K - k * BLOCK_SIZE_K), + other=0.0, + ) + b = tl.load(b_ptrs, mask=offs_k[:, None] < K - k * BLOCK_SIZE_K, other=0.0) + # We accumulate along the K dimension. + if use_int8_w8a16: + accumulator = tl.dot(a, b.to(compute_type), acc=accumulator) + elif use_fp8_w8a8: + if group_k > 0 and group_n > 0: + k_start = k * BLOCK_SIZE_K + offs_ks = k_start // group_k + a_scale = tl.load( + a_scale_ptrs + offs_ks * stride_ask, mask=token_mask, other=0.0 + ) + b_scale = tl.load(b_scale_ptrs + offs_ks * stride_bsk) + + accumulator += tl.dot(a, b) * a_scale[:, None] * b_scale[None, :] + else: + accumulator = tl.dot(a, b, acc=accumulator) + else: + accumulator += tl.dot(a, b) + # Advance the ptrs to the next K block. + a_ptrs += BLOCK_SIZE_K * stride_ak + b_ptrs += BLOCK_SIZE_K * stride_bk + + if MUL_ROUTED_WEIGHT: + moe_weight = tl.load(topk_weights_ptr + offs_token, mask=token_mask, other=0) + accumulator = accumulator * moe_weight[:, None] + if use_int8_w8a16: + accumulator = (accumulator * b_scale).to(compute_type) + elif use_fp8_w8a8: + if group_k > 0 and group_n > 0: + accumulator = accumulator.to(compute_type) + else: + accumulator = (accumulator * a_scale * b_scale).to(compute_type) + else: + accumulator = accumulator.to(compute_type) + # ----------------------------------------------------------- + # Write back the block of the output + offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N) + c_ptrs = c_ptr + stride_cm * offs_token[:, None] + stride_cn * offs_cn[None, :] + c_mask = token_mask[:, None] & (offs_cn[None, :] < N) + tl.store(c_ptrs, accumulator, mask=c_mask) + + +def invoke_fused_moe_kernel( + A: torch.Tensor, + B: torch.Tensor, + C: torch.Tensor, + A_scale: Optional[torch.Tensor], + B_scale: Optional[torch.Tensor], + B_zp: Optional[torch.Tensor], + topk_weights: Optional[torch.Tensor], + sorted_token_ids: torch.Tensor, + expert_ids: torch.Tensor, + num_tokens_post_padded: torch.Tensor, + mul_routed_weight: bool, + top_k: int, + config: Dict[str, Any], + compute_type: tl.dtype, + use_fp8_w8a8: bool, + use_int8_w8a16: bool, + use_int4_w4a16: bool, + block_shape: Optional[List[int]] = None, +) -> None: + assert topk_weights is not None or not mul_routed_weight + assert topk_weights is None or topk_weights.stride(1) == 1 + assert sorted_token_ids.stride(0) == 1 + + if use_fp8_w8a8: + assert B_scale is not None + assert ( + block_shape is None + or triton.cdiv(B.shape[-2], block_shape[0]) == B_scale.shape[-2] + ) + assert ( + block_shape is None + or triton.cdiv(B.shape[-1], block_shape[1]) == B_scale.shape[-1] + ) + + elif use_int8_w8a16 or use_int4_w4a16: + assert B_scale is not None + assert block_shape is None or block_shape[0] == 0 + else: + assert A_scale is None + assert B_scale is None + + M = A.shape[0] + num_tokens = M * top_k + + EM = sorted_token_ids.shape[0] + if A.shape[0] < config["BLOCK_SIZE_M"]: + # optimize for small batch_size. + # We assume that top_ids of each token is unique, so + # so num_valid_experts <= batch_size <= BLOCK_SIZE_M, + # and we can skip some invalid blocks. + EM = min(sorted_token_ids.shape[0], A.shape[0] * top_k * config["BLOCK_SIZE_M"]) + grid = lambda META: ( + triton.cdiv(EM, META["BLOCK_SIZE_M"]) + * triton.cdiv(B.shape[1], META["BLOCK_SIZE_N"]), + ) + + if ( + (use_int8_w8a16 or use_int4_w4a16) + and block_shape is not None + and block_shape[1] > 0 + ): + assert B_scale is not None and B_scale.ndim == 3 + assert B_zp is None or B_zp.ndim == 3 + + use_moe_wna16_cuda = should_moe_wna16_use_cuda( + num_valid_tokens=num_tokens, + group_size=block_shape[1], + num_experts=B.shape[0], + bit=4 if use_int4_w4a16 else 8, + ) + config = config.copy() + config.update( + get_moe_wna16_block_config( + config=config, + use_moe_wna16_cuda=use_moe_wna16_cuda, + num_valid_tokens=num_tokens, + size_k=A.shape[1], + size_n=B.shape[1], + num_experts=B.shape[1], + group_size=block_shape[1], + real_top_k=top_k, + block_size_m=config["BLOCK_SIZE_M"], + ) + ) + + if use_moe_wna16_cuda: + bit = 4 if use_int4_w4a16 else 8 + ops.moe_wna16_gemm( + A, + C, + B, + B_scale, + B_zp, + topk_weights if mul_routed_weight else None, + sorted_token_ids, + expert_ids, + num_tokens_post_padded, + top_k, + config["BLOCK_SIZE_M"], + config["BLOCK_SIZE_N"], + config["BLOCK_SIZE_K"], + bit, + ) + return + + fused_moe_kernel_gptq_awq[grid]( + A, + B, + C, + B_scale, + B_zp, + topk_weights, + sorted_token_ids, + expert_ids, + num_tokens_post_padded, + B.shape[1], + A.shape[1], + EM, + num_tokens, + A.stride(0), + A.stride(1), + B.stride(0), + B.stride(2), + B.stride(1), + C.stride(1), + C.stride(2), + B_scale.stride(0), + B_scale.stride(2), + B_scale.stride(1), + B_zp.stride(0) if B_zp is not None else 0, + B_zp.stride(2) if B_zp is not None else 0, + B_zp.stride(1) if B_zp is not None else 0, + block_k_diviable=A.shape[1] % config["BLOCK_SIZE_K"] == 0, + group_size=block_shape[1], + MUL_ROUTED_WEIGHT=mul_routed_weight, + top_k=top_k, + compute_type=compute_type, + has_zp=B_zp is not None, + use_int4_w4a16=use_int4_w4a16, + use_int8_w8a16=use_int8_w8a16, + **config, + ) + else: + config = config.copy() + BLOCK_SIZE_K = config.pop("BLOCK_SIZE_K") + if block_shape is not None: + BLOCK_SIZE_K = min(BLOCK_SIZE_K, min(block_shape[0], block_shape[1])) + fused_moe_kernel[grid]( + A, + B, + C, + A_scale, + B_scale, + topk_weights, + sorted_token_ids, + expert_ids, + num_tokens_post_padded, + B.shape[1], + B.shape[2], + EM, + num_tokens, + A.stride(0), + A.stride(1), + B.stride(0), + B.stride(2), + B.stride(1), + C.stride(1), + C.stride(2), + A_scale.stride(0) if A_scale is not None and A_scale.ndim == 2 else 0, + A_scale.stride(1) if A_scale is not None and A_scale.ndim == 2 else 0, + B_scale.stride(0) if B_scale is not None and B_scale.ndim >= 2 else 0, + B_scale.stride(2) if B_scale is not None and B_scale.ndim == 3 else 0, + B_scale.stride(1) if B_scale is not None and B_scale.ndim >= 2 else 0, + 0 if block_shape is None else block_shape[0], + 0 if block_shape is None else block_shape[1], + MUL_ROUTED_WEIGHT=mul_routed_weight, + top_k=top_k, + compute_type=compute_type, + use_fp8_w8a8=use_fp8_w8a8, + use_int8_w8a16=use_int8_w8a16, + BLOCK_SIZE_K=BLOCK_SIZE_K, + **config, + ) + + +# Adapted from: https://github.com/sgl-project/sglang/pull/2628 +def get_config_file_name( + E: int, N: int, dtype: Optional[str], block_shape: Optional[List[int]] = None +) -> str: + device_name = current_platform.get_device_name().replace(" ", "_") + dtype_selector = "" if not dtype else f",dtype={dtype}" + block_shape_selector = ( + "" if not block_shape or not all(block_shape) else f",block_shape={block_shape}" + ).replace(" ", "") + return f"E={E},N={N},device_name={device_name}{dtype_selector}{block_shape_selector}.json" # noqa: E501 + + +# Adapted from: https://github.com/sgl-project/sglang/pull/2628 +@functools.lru_cache +def get_moe_configs( + E: int, + N: int, + dtype: Optional[str], + block_n: Optional[int] = None, + block_k: Optional[int] = None, +) -> Optional[Dict[int, Any]]: + """ + Return optimized configurations for the fused MoE kernel. + + The return value will be a dictionary that maps an irregular grid of + batch sizes to configurations of the fused_moe kernel. To evaluate the + kernel on a given batch size bs, the closest batch size in the grid should + be picked and the associated configuration chosen to invoke the kernel. + """ + + # First look up if an optimized configuration is available in the configs + # directory + block_shape = [block_n, block_k] if block_n and block_k else None + json_file_name = get_config_file_name(E, N, dtype, block_shape) + + config_file_path = os.path.join( + os.path.dirname(os.path.realpath(__file__)), "configs", json_file_name + ) + if os.path.exists(config_file_path): + with open(config_file_path) as f: + logger.info("Using configuration from %s for MoE layer.", config_file_path) + # If a configuration has been found, return it + return {int(key): val for key, val in json.load(f).items()} + + # If no optimized configuration is available, we will use the default + # configuration + logger.warning( + ( + "Using default MoE config. Performance might be sub-optimal! " + "Config file not found at %s" + ), + config_file_path, + ) + return None + + +def get_moe_wna16_block_config( + config: Dict[str, int], + use_moe_wna16_cuda: bool, + num_valid_tokens: int, + size_k: int, + size_n: int, + num_experts: int, + group_size: int, + real_top_k: int, + block_size_m: int, +): + if "BLOCK_SIZE_N" in config and "BLOCK_SIZE_K" in config: + # optimal block config is set + return {} + if not use_moe_wna16_cuda: + # triton moe wna16 kernel + if num_valid_tokens // real_top_k == 1: + # if bs=1, use a smaller BLOCK_SIZE_N + return {"BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64} + else: + return {"BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32} + else: + # cuda moe wna16 kernel + # set default block_size 128, and increase them when num_blocks + # is too large. + block_size_n = 128 + block_size_k = 128 + if block_size_k <= group_size: + block_size_k = group_size + + num_n_blocks = size_k // block_size_k + num_k_blocks = size_n // block_size_k + num_m_blocks = ( + num_valid_tokens + block_size_m - 1 + ) / block_size_m + num_experts + if num_valid_tokens // real_top_k <= block_size_m: + num_m_blocks = min(num_m_blocks, num_valid_tokens) + num_blocks = num_m_blocks * num_n_blocks * num_k_blocks + + if size_k % 256 == 0 and num_blocks >= 256 and block_size_k < 256: + block_size_k = 256 + num_blocks = num_blocks // (256 // block_size_k) + + if ( + num_m_blocks <= 16 + and size_k % (block_size_k * 2) == 0 + and size_k % (block_size_k * 2) == 0 + and block_size_k <= 512 + and num_blocks >= 512 + ): + block_size_k = block_size_k * 2 + num_blocks = num_blocks // 2 + + if num_blocks > 1024: + block_size_n = 256 + num_n_blocks = num_n_blocks // 2 + num_blocks = num_blocks // 2 + + if size_n <= 1024 and num_blocks >= 1024: + # The kernel performance got much better with BLOCK_SIZE_N=1024 + # when num_blocks is large, event when N is small. + # Not sure why, maybe it force the CUDA SM process only one block + # at the same time. + block_size_n = 1024 + + return {"BLOCK_SIZE_N": block_size_n, "BLOCK_SIZE_K": block_size_k} + + +def should_moe_wna16_use_cuda( + num_valid_tokens: int, group_size: int, num_experts: int, bit: int +): + return ( + bit == 4 and group_size in [32, 64, 128] and num_valid_tokens / num_experts <= 6 + ) + + +def get_default_config( + M: int, + E: int, + N: int, + K: int, + topk: int, + dtype: Optional[str], + is_marlin: bool, + block_shape: Optional[List[int]] = None, +) -> Dict[str, int]: + if dtype == "fp8_w8a8" and block_shape is not None: + # Block-wise quant: BLOCK_SIZE_N must be divisible by block_shape[0] + # BLOCK_SIZE_K must be divisible by block_shape[1] + config = { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": block_shape[0], + "BLOCK_SIZE_K": block_shape[1], + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3, + } + elif dtype in ["int4_w4a16", "int8_w8a16"] and block_shape is not None: + # moe wna16 kernels + # only set BLOCK_SIZE_M + # BLOCK_SIZE_N and BLOCK_SIZE_K would be set later + bit = 4 if dtype == "int4_w4a16" else 8 + use_moe_wna16_cuda = should_moe_wna16_use_cuda(M * topk, block_shape[1], E, bit) + if use_moe_wna16_cuda: + config = {"BLOCK_SIZE_M": min(16, M)} + elif M <= 20: + config = {"BLOCK_SIZE_M": 16, "GROUP_SIZE_M": 1} + elif M <= 40: + config = {"BLOCK_SIZE_M": 32, "GROUP_SIZE_M": 1} + else: + config = {"BLOCK_SIZE_M": 64, "GROUP_SIZE_M": 1} + else: + config = { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 8, + } + # A heuristic: fused marlin works faster with this config for small M + if M <= E or (is_marlin and M <= 32): + config = { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + } + return config + + +def try_get_optimal_moe_config( + w1_shape: Tuple[int, ...], + w2_shape: Tuple[int, ...], + top_k: int, + dtype: Optional[str], + M: int, + is_marlin: bool = False, + block_shape: Optional[List[int]] = None, +): + # from vllm.model_executor.layers.fused_moe import get_config + # TODO: removed when syncing to vLLM, do we need this? + # override_config = get_config() + override_config = None + if override_config: + config = override_config + else: + # First try to load optimal config from the file + E, _, N = w2_shape + if dtype == "int4_w4a16": + N = N * 2 + block_n = block_shape[0] if block_shape else 0 + block_k = block_shape[1] if block_shape else 0 + configs = get_moe_configs(E, N, dtype, block_n, block_k) + + if configs: + # If an optimal configuration map has been found, look up the + # optimal config + config = configs[min(configs.keys(), key=lambda x: abs(x - M))] + else: + # Else use the default config + config = get_default_config( + M, E, N, w1_shape[2], top_k, dtype, is_marlin, block_shape + ) + return config + + +def vllm_topk_softmax( + topk_weights: torch.Tensor, + topk_indices: torch.Tensor, + token_expert_indices: torch.Tensor, + gating_output: torch.Tensor, + renormalize: bool, +) -> tuple[torch.Tensor, ...]: + ops.topk_softmax( + topk_weights, + topk_indices, + token_expert_indices, + gating_output, + ) + if renormalize: + topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True) + + return topk_weights, topk_indices + + +def dispatch_topk_func() -> Callable[..., tuple[torch.Tensor, ...]]: + return vllm_topk_softmax + + +def fused_topk( + hidden_states: torch.Tensor, + gating_output: torch.Tensor, + topk: int, + renormalize: bool, +) -> Tuple[torch.Tensor, torch.Tensor]: + assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch" + + M, _ = hidden_states.shape + + topk_weights = torch.empty( + M, topk, dtype=torch.float32, device=hidden_states.device + ) + topk_ids = torch.empty(M, topk, dtype=torch.int32, device=hidden_states.device) + token_expert_indicies = torch.empty( + M, topk, dtype=torch.int32, device=hidden_states.device + ) + + gating_output_float = gating_output.float() # TODO(woosuk): Optimize this. + + topk_func = dispatch_topk_func() + topk_weights, topk_ids = topk_func( + topk_weights, topk_ids, token_expert_indicies, gating_output_float, renormalize + ) + + del token_expert_indicies # Not used. Will be used in the future. + return topk_weights, topk_ids + + +# This is used by the Deepseek-V2 and Deepseek-V3 model +@torch.compile(dynamic=True, backend=current_platform.simple_compile_backend) +def grouped_topk( + hidden_states: torch.Tensor, + gating_output: torch.Tensor, + topk: int, + renormalize: bool, + num_expert_group: int = 0, + topk_group: int = 0, + scoring_func: str = "softmax", + e_score_correction_bias: Optional[torch.Tensor] = None, +) -> Tuple[torch.Tensor, torch.Tensor]: + + assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch" + + if scoring_func == "softmax": + scores = torch.softmax(gating_output, dim=-1) + elif scoring_func == "sigmoid": + scores = gating_output.sigmoid() + else: + raise ValueError(f"Unsupported scoring function: {scoring_func}") + + num_token = scores.shape[0] + if e_score_correction_bias is not None: + # Store original scores before applying correction bias. We use biased + # scores for expert selection but original scores for routing weights + original_scores = scores + scores = scores + e_score_correction_bias.unsqueeze(0) + group_scores = ( + scores.view(num_token, num_expert_group, -1).topk(2, dim=-1)[0].sum(dim=-1) + ) + else: + group_scores = ( + scores.view(num_token, num_expert_group, -1).max(dim=-1).values + ) # [n, n_group] + group_idx = torch.topk(group_scores, k=topk_group, dim=-1, sorted=False)[ + 1 + ] # [n, top_k_group] + group_mask = torch.zeros_like(group_scores) # [n, n_group] + group_mask.scatter_(1, group_idx, 1) # [n, n_group] + score_mask = ( + group_mask.unsqueeze(-1) + .expand(num_token, num_expert_group, scores.shape[-1] // num_expert_group) + .reshape(num_token, -1) + ) # [n, e] + tmp_scores = scores.masked_fill(~score_mask.bool(), float("-inf")) # [n, e] + + if e_score_correction_bias is not None: + topk_ids = torch.topk(tmp_scores, k=topk, dim=-1, sorted=False)[1] + # Use original unbiased scores for the routing weights + topk_weights = original_scores.gather(1, topk_ids) + else: + topk_weights, topk_ids = torch.topk(tmp_scores, k=topk, dim=-1, sorted=False) + + if renormalize: + topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True) + + return topk_weights.to(torch.float32), topk_ids.to(torch.int32) + + +def get_config_dtype_str( + dtype: torch.dtype, + use_int4_w4a16: Optional[bool] = False, + use_int8_w8a16: Optional[bool] = False, + use_fp8_w8a8: Optional[bool] = False, +) -> Optional[str]: + if use_fp8_w8a8: + return "fp8_w8a8" + elif use_int8_w8a16: + return "int8_w8a16" + elif use_int4_w4a16: + return "int4_w4a16" + elif dtype == torch.float: + # avoiding cases where kernel fails when float32 MoE + # use fp16/bfloat16 configs + return "float32" + return None + + +def inplace_fused_experts( + hidden_states: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + activation: str = "silu", + apply_router_weight_on_input: bool = False, + use_fp8_w8a8: bool = False, + use_int8_w8a16: bool = False, + use_int4_w4a16: bool = False, + global_num_experts: int = -1, + expert_map: Optional[torch.Tensor] = None, + w1_scale: Optional[torch.Tensor] = None, + w2_scale: Optional[torch.Tensor] = None, + w1_zp: Optional[torch.Tensor] = None, + w2_zp: Optional[torch.Tensor] = None, + a1_scale: Optional[torch.Tensor] = None, + a2_scale: Optional[torch.Tensor] = None, + block_shape: Optional[List[int]] = None, +) -> None: + fused_experts_impl( + hidden_states, + w1, + w2, + topk_weights, + topk_ids, + True, + activation, + apply_router_weight_on_input, + use_fp8_w8a8, + use_int8_w8a16, + use_int4_w4a16, + global_num_experts, + expert_map, + w1_scale, + w2_scale, + w1_zp, + w2_zp, + a1_scale, + a2_scale, + block_shape, + ) + + +def inplace_fused_experts_fake( + hidden_states: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + activation: str = "silu", + apply_router_weight_on_input: bool = False, + use_fp8_w8a8: bool = False, + use_int8_w8a16: bool = False, + use_int4_w4a16: bool = False, + global_num_experts: int = -1, + expert_map: Optional[torch.Tensor] = None, + w1_scale: Optional[torch.Tensor] = None, + w2_scale: Optional[torch.Tensor] = None, + w1_zp: Optional[torch.Tensor] = None, + w2_zp: Optional[torch.Tensor] = None, + a1_scale: Optional[torch.Tensor] = None, + a2_scale: Optional[torch.Tensor] = None, + block_shape: Optional[List[int]] = None, +) -> None: + pass + + +direct_register_custom_op( + op_name="inplace_fused_experts", + op_func=inplace_fused_experts, + mutates_args=["hidden_states"], + fake_impl=inplace_fused_experts_fake, +) + + +def outplace_fused_experts( + hidden_states: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + activation: str = "silu", + apply_router_weight_on_input: bool = False, + use_fp8_w8a8: bool = False, + use_int8_w8a16: bool = False, + use_int4_w4a16: bool = False, + global_num_experts: int = -1, + expert_map: Optional[torch.Tensor] = None, + w1_scale: Optional[torch.Tensor] = None, + w2_scale: Optional[torch.Tensor] = None, + w1_zp: Optional[torch.Tensor] = None, + w2_zp: Optional[torch.Tensor] = None, + a1_scale: Optional[torch.Tensor] = None, + a2_scale: Optional[torch.Tensor] = None, + block_shape: Optional[List[int]] = None, +) -> torch.Tensor: + return fused_experts_impl( + hidden_states, + w1, + w2, + topk_weights, + topk_ids, + False, + activation, + apply_router_weight_on_input, + use_fp8_w8a8, + use_int8_w8a16, + use_int4_w4a16, + global_num_experts, + expert_map, + w1_scale, + w2_scale, + w1_zp, + w2_zp, + a1_scale, + a2_scale, + block_shape, + ) + + +def outplace_fused_experts_fake( + hidden_states: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + activation: str = "silu", + apply_router_weight_on_input: bool = False, + use_fp8_w8a8: bool = False, + use_int8_w8a16: bool = False, + use_int4_w4a16: bool = False, + global_num_experts: int = -1, + expert_map: Optional[torch.Tensor] = None, + w1_scale: Optional[torch.Tensor] = None, + w2_scale: Optional[torch.Tensor] = None, + w1_zp: Optional[torch.Tensor] = None, + w2_zp: Optional[torch.Tensor] = None, + a1_scale: Optional[torch.Tensor] = None, + a2_scale: Optional[torch.Tensor] = None, + block_shape: Optional[List[int]] = None, +) -> torch.Tensor: + return torch.empty_like(hidden_states) + + +direct_register_custom_op( + op_name="outplace_fused_experts", + op_func=outplace_fused_experts, + mutates_args=[], + fake_impl=outplace_fused_experts_fake, +) + + +def torch_vllm_inplace_fused_experts(**kwargs) -> torch.Tensor: + ops.inplace_fused_experts(**kwargs) + hidden_states = kwargs["hidden_states"] + return hidden_states + + +def torch_vllm_outplace_fused_experts(**kwargs) -> torch.Tensor: + return ops.outplace_fused_experts(**kwargs) + + +def dispatch_fused_experts_func(inplace: bool) -> Callable[..., torch.Tensor]: + if inplace: + return torch_vllm_inplace_fused_experts + return torch_vllm_outplace_fused_experts + + +def fused_experts( + hidden_states: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + inplace: bool = False, + activation: str = "silu", + apply_router_weight_on_input: bool = False, + use_fp8_w8a8: bool = False, + use_int8_w8a16: bool = False, + use_int4_w4a16: bool = False, + global_num_experts: int = -1, + expert_map: Optional[torch.Tensor] = None, + w1_scale: Optional[torch.Tensor] = None, + w2_scale: Optional[torch.Tensor] = None, + w1_zp: Optional[torch.Tensor] = None, + w2_zp: Optional[torch.Tensor] = None, + a1_scale: Optional[torch.Tensor] = None, + a2_scale: Optional[torch.Tensor] = None, + block_shape: Optional[List[int]] = None, +) -> torch.Tensor: + return dispatch_fused_experts_func(inplace)( + hidden_states=hidden_states, + w1=w1, + w2=w2, + topk_weights=topk_weights, + topk_ids=topk_ids, + activation=activation, + apply_router_weight_on_input=apply_router_weight_on_input, + use_fp8_w8a8=use_fp8_w8a8, + use_int8_w8a16=use_int8_w8a16, + use_int4_w4a16=use_int4_w4a16, + global_num_experts=global_num_experts, + expert_map=expert_map, + w1_scale=w1_scale, + w2_scale=w2_scale, + w1_zp=w1_zp, + w2_zp=w2_zp, + a1_scale=a1_scale, + a2_scale=a2_scale, + block_shape=block_shape, + ) + + +def fused_experts_impl( + hidden_states: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + inplace: bool = False, + activation: str = "silu", + apply_router_weight_on_input: bool = False, + use_fp8_w8a8: bool = False, + use_int8_w8a16: bool = False, + use_int4_w4a16: bool = False, + global_num_experts: int = -1, + expert_map: Optional[torch.Tensor] = None, + w1_scale: Optional[torch.Tensor] = None, + w2_scale: Optional[torch.Tensor] = None, + w1_zp: Optional[torch.Tensor] = None, + w2_zp: Optional[torch.Tensor] = None, + a1_scale: Optional[torch.Tensor] = None, + a2_scale: Optional[torch.Tensor] = None, + block_shape: Optional[List[int]] = None, +): + # Check constraints. + if use_int4_w4a16: + assert hidden_states.shape[1] // 2 == w1.shape[2], "Hidden size mismatch" + else: + assert hidden_states.shape[1] == w1.shape[2], "Hidden size mismatch" + + assert topk_weights.shape == topk_ids.shape, "topk shape mismatch" + assert hidden_states.is_contiguous(), "Hidden_states must be contiguous" + assert w1.stride(-1) == 1, "Stride of last dimension must be 1" + assert w2.stride(-1) == 1, "Stride of last dimension must be 1" + assert hidden_states.dtype in [torch.float32, torch.float16, torch.bfloat16] + + num_tokens, _ = hidden_states.shape + E, N, _ = w1.shape + K = w2.shape[1] + if global_num_experts == -1: + global_num_experts = E + top_k_num = topk_ids.shape[1] + # We execute the fused_moe kernel in chunks to circumvent this issue: + # https://github.com/vllm-project/vllm/issues/5938 + CHUNK_SIZE = VLLM_FUSED_MOE_CHUNK_SIZE + M = min(num_tokens, CHUNK_SIZE) + config_dtype = get_config_dtype_str( + use_fp8_w8a8=use_fp8_w8a8, + use_int8_w8a16=use_int8_w8a16, + use_int4_w4a16=use_int4_w4a16, + dtype=hidden_states.dtype, + ) + + get_config_func = functools.partial( + try_get_optimal_moe_config, + w1.shape, + w2.shape, + top_k_num, + config_dtype, + block_shape=block_shape, + ) + + config = get_config_func(M) + + # We can reuse the memory between these because by the time we need + # cache3, we're done with cache1 + cache13 = torch.empty( + M * top_k_num * max(N, K), + device=hidden_states.device, + dtype=hidden_states.dtype, + ) + intermediate_cache1 = cache13[: M * top_k_num * N].view(M, top_k_num, N) + intermediate_cache3 = cache13[: M * top_k_num * K].view(M, top_k_num, K) + + # This needs separate memory since it's used concurrently with cache1 + intermediate_cache2 = torch.empty( + (M * top_k_num, N // 2), device=hidden_states.device, dtype=hidden_states.dtype + ) + + if hidden_states.dtype == torch.bfloat16: + compute_type = tl.bfloat16 + elif hidden_states.dtype == torch.float16: + compute_type = tl.float16 + elif hidden_states.dtype == torch.float32: + compute_type = tl.float32 + else: + raise ValueError(f"Unsupported compute_type: {hidden_states.dtype}") + + if inplace: + out_hidden_states = hidden_states + else: + out_hidden_states = torch.empty_like(hidden_states) + + for chunk in range((num_tokens // CHUNK_SIZE) + 1): + begin_chunk_idx, end_chunk_idx = ( + chunk * CHUNK_SIZE, + min((chunk + 1) * CHUNK_SIZE, num_tokens), + ) + curr_hidden_states = hidden_states[begin_chunk_idx:end_chunk_idx] + tokens_in_chunk, _ = curr_hidden_states.shape + + if tokens_in_chunk == 0: + break + + if tokens_in_chunk < CHUNK_SIZE and chunk > 0: + # Adjust the intermediate cache size and config for the last + # chunk. Note that in most cases we only have one chunk + # so the cache size and config are already set correctly and + # do not need to be adjusted. + intermediate_cache1 = intermediate_cache1[:tokens_in_chunk] + intermediate_cache2 = intermediate_cache2[ + : tokens_in_chunk * topk_ids.shape[1] + ] + intermediate_cache3 = intermediate_cache3[:tokens_in_chunk] + config = get_config_func(tokens_in_chunk) + + curr_topk_ids = topk_ids[begin_chunk_idx:end_chunk_idx] + curr_topk_weights = topk_weights[begin_chunk_idx:end_chunk_idx] + + a1q_scale: Optional[torch.Tensor] = None + + if use_fp8_w8a8: + qcurr_hidden_states, a1q_scale = _fp8_quantize( + curr_hidden_states, a1_scale, block_shape + ) + else: + qcurr_hidden_states = curr_hidden_states + a1q_scale = a1_scale + + sorted_token_ids, expert_ids, num_tokens_post_padded = moe_align_block_size( + curr_topk_ids, config["BLOCK_SIZE_M"], global_num_experts, expert_map + ) + + invoke_fused_moe_kernel( + qcurr_hidden_states, + w1, + intermediate_cache1, + a1q_scale, + w1_scale, + w1_zp, + curr_topk_weights, + sorted_token_ids, + expert_ids, + num_tokens_post_padded, + apply_router_weight_on_input, + top_k_num, + config, + compute_type=compute_type, + use_fp8_w8a8=use_fp8_w8a8, + use_int8_w8a16=use_int8_w8a16, + use_int4_w4a16=use_int4_w4a16, + block_shape=block_shape, + ) + + if activation == "silu": + ops.silu_and_mul(intermediate_cache2, intermediate_cache1.view(-1, N)) + elif activation == "gelu": + ops.gelu_and_mul(intermediate_cache2, intermediate_cache1.view(-1, N)) + else: + raise ValueError(f"Unsupported FusedMoe activation: {activation}") + + a2q_scale: Optional[torch.Tensor] = None + + if use_fp8_w8a8: + qintermediate_cache2, a2q_scale = _fp8_quantize( + intermediate_cache2, a2_scale, block_shape + ) + else: + qintermediate_cache2 = intermediate_cache2 + a2q_scale = a2_scale + + invoke_fused_moe_kernel( + qintermediate_cache2, + w2, + intermediate_cache3, + a2q_scale, + w2_scale, + w2_zp, + curr_topk_weights, + sorted_token_ids, + expert_ids, + num_tokens_post_padded, + not apply_router_weight_on_input, + 1, + config, + compute_type=compute_type, + use_fp8_w8a8=use_fp8_w8a8, + use_int8_w8a16=use_int8_w8a16, + use_int4_w4a16=use_int4_w4a16, + block_shape=block_shape, + ) + + ops.moe_sum( + intermediate_cache3.view(*intermediate_cache3.shape), + out_hidden_states[begin_chunk_idx:end_chunk_idx], + ) + + return out_hidden_states + + +def fused_moe( + hidden_states: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + gating_output: torch.Tensor, + topk: int, + renormalize: bool, + inplace: bool = False, + activation: str = "silu", + apply_router_weight_on_input: bool = False, + use_grouped_topk: bool = False, + num_expert_group: Optional[int] = None, + topk_group: Optional[int] = None, + custom_routing_function: Optional[Callable] = None, + use_fp8_w8a8: bool = False, + use_int8_w8a16: bool = False, + use_int4_w4a16: bool = False, + global_num_experts: int = -1, + expert_map: Optional[torch.Tensor] = None, + w1_scale: Optional[torch.Tensor] = None, + w2_scale: Optional[torch.Tensor] = None, + w1_zp: Optional[torch.Tensor] = None, + w2_zp: Optional[torch.Tensor] = None, + a1_scale: Optional[torch.Tensor] = None, + a2_scale: Optional[torch.Tensor] = None, + block_shape: Optional[List[int]] = None, +) -> torch.Tensor: + """ + This function computes a Mixture of Experts (MoE) layer using two sets of + weights, w1 and w2, and top-k gating mechanism. + + Parameters: + - hidden_states (torch.Tensor): The input tensor to the MoE layer. + - w1 (torch.Tensor): The first set of expert weights. + - w2 (torch.Tensor): The second set of expert weights. + - gating_output (torch.Tensor): The output of the gating operation + (before softmax). + - topk (int): The number of top-k experts to select. + - renormalize (bool): If True, renormalize the top-k weights to sum to 1. + - inplace (bool): If True, perform the operation in-place. + Defaults to False. + - activation (str): The activation function to apply after the first + MoE layer. + - num_expert_group: Optional[int]: additional parameter for grouped_topk + - topk_group: Optional[int]: additional parameter for grouped_topk + - use_grouped_topk: If True, use grouped_topk instead of fused_topk + note: Deepseekv2 model uses grouped_topk + - use_fp8_w8a8 (bool): If True, use fp8 arithmetic to compute the inner + products for w1 and w2. Defaults to False. + - use_int8_w8a16 (bool): If True, use matmul of int8 weight and bf16/fp16 + activation to compute the inner products for w1 and w2. + Defaults to False. + - use_int4_w4a16 (bool): If True, use matmul of int4 weight and bf16/fp16 + activation to compute the inner products for w1 and w2. + Defaults to False. + - global_num_experts (int): The total number of experts in the global + expert space. + - expert_map (Optional[torch.Tensor]): A tensor mapping expert indices + from the global expert space to the local expert space of the expert + parallel shard. + - w1_scale (Optional[torch.Tensor]): Optional scale to be used for + w1. + - w2_scale (Optional[torch.Tensor]): Optional scale to be used for + w2. + - a1_scale (Optional[torch.Tensor]): Optional scale to be used for + a1. + - a2_scale (Optional[torch.Tensor]): Optional scale to be used for + a2. + - block_shape: (Optional[List[int]]): Optional block size for block-wise + quantization. + + Returns: + - torch.Tensor: The output tensor after applying the MoE layer. + """ + + if use_grouped_topk: + assert num_expert_group is not None and topk_group is not None + topk_weights, topk_ids = grouped_topk( + hidden_states, + gating_output, + topk, + renormalize, + num_expert_group, + topk_group, + ) + elif custom_routing_function is None: + topk_weights, topk_ids = fused_topk( + hidden_states, gating_output, topk, renormalize + ) + else: + topk_weights, topk_ids = custom_routing_function( + hidden_states, gating_output, topk, renormalize + ) + + return fused_experts( + hidden_states, + w1, + w2, + topk_weights, + topk_ids, + inplace=inplace, + activation=activation, + apply_router_weight_on_input=apply_router_weight_on_input, + use_fp8_w8a8=use_fp8_w8a8, + use_int8_w8a16=use_int8_w8a16, + use_int4_w4a16=use_int4_w4a16, + global_num_experts=global_num_experts, + expert_map=expert_map, + w1_scale=w1_scale, + w2_scale=w2_scale, + w1_zp=w1_zp, + w2_zp=w2_zp, + a1_scale=a1_scale, + a2_scale=a2_scale, + block_shape=block_shape, + ) diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/layers.py b/build/torch26-cxx98-cu126-aarch64-linux/moe/layers.py new file mode 100644 index 0000000000000000000000000000000000000000..1befea3f9e97a6a4aa1679821d30091c3d825bb8 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/layers.py @@ -0,0 +1,64 @@ +import torch +import torch.nn as nn + +from .fused_moe import fused_moe + + +def _llama4_topk( + hidden_states: torch.Tensor, + gating_output: torch.Tensor, + topk: int, + renormalize: bool, +): + assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch" + topk_weights, topk_ids = torch.topk(gating_output, topk, dim=1) + topk_weights = torch.sigmoid(topk_weights.float()).to(hidden_states.dtype) + return topk_weights, topk_ids + + +def _fix_llama4_experts(hidden_states: torch.Tensor, experts: nn.Module): + if experts.gate_up_proj.shape[-1] != hidden_states.shape[-1]: + experts.gate_up_proj = nn.Parameter( + experts.gate_up_proj.transpose(1, 2).contiguous() + ) + experts.down_proj = nn.Parameter(experts.down_proj.transpose(1, 2).contiguous()) + + +class Llama4TextMoe(nn.Module): + has_backward = False + + experts: nn.Module + router: nn.Linear + shared_expert: nn.Module + top_k: int + + def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: + hidden_states = hidden_states.reshape(-1, hidden_states.shape[-1]) + + _fix_llama4_experts(hidden_states, self.experts) + + router_logits = self.router(hidden_states) + + extra_kwargs = {} + use_fp8_w8a8 = False + if hasattr(self.experts, "gate_up_proj_scale"): + use_fp8_w8a8 = True + extra_kwargs["w1_scale"] = self.experts.gate_up_proj_scale + extra_kwargs["w2_scale"] = self.experts.down_proj_scale + + out = fused_moe( + hidden_states, + w1=self.experts.gate_up_proj, + w2=self.experts.down_proj, + gating_output=router_logits, + topk=self.top_k, + renormalize=False, + custom_routing_function=_llama4_topk, + apply_router_weight_on_input=True, + use_fp8_w8a8=use_fp8_w8a8, + **extra_kwargs + ) + + out += self.shared_expert(hidden_states) + + return out, router_logits.t() diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/moe_align_block_size.py b/build/torch26-cxx98-cu126-aarch64-linux/moe/moe_align_block_size.py new file mode 100644 index 0000000000000000000000000000000000000000..a698de5aaa123df5a2d663f252e1d5914e880d09 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/moe_align_block_size.py @@ -0,0 +1,252 @@ +# SPDX-License-Identifier: Apache-2.0 +import os +from typing import Optional, Tuple + +import torch +import triton +import triton.language as tl + +from ._ops import ops + + +def round_up(x: int, y: int) -> int: + return ((x + y - 1) // y) * y + + +VLLM_ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON = bool( + int(os.getenv("VLLM_ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON", "0")) +) + + +def ceil_div(a, b): + return (a + b - 1) // b + + +@triton.jit +def moe_align_block_size_stage1( + topk_ids_ptr, + tokens_cnts_ptr, + num_experts: tl.constexpr, + numel: tl.constexpr, + tokens_per_thread: tl.constexpr, +): + pid = tl.program_id(0) + + start_idx = pid * tokens_per_thread + + off_c = (pid + 1) * num_experts + + for i in range(tokens_per_thread): + if start_idx + i < numel: + idx = tl.load(topk_ids_ptr + start_idx + i) + token_cnt = tl.load(tokens_cnts_ptr + off_c + idx) + tl.store(tokens_cnts_ptr + off_c + idx, token_cnt + 1) + + +@triton.jit +def moe_align_block_size_stage2( + tokens_cnts_ptr, + num_experts: tl.constexpr, +): + pid = tl.program_id(0) + + last_cnt = 0 + for i in range(1, num_experts + 1): + token_cnt = tl.load(tokens_cnts_ptr + i * num_experts + pid) + last_cnt = last_cnt + token_cnt + tl.store(tokens_cnts_ptr + i * num_experts + pid, last_cnt) + + +@triton.jit +def moe_align_block_size_stage3( + total_tokens_post_pad_ptr, + tokens_cnts_ptr, + cumsum_ptr, + num_experts: tl.constexpr, + block_size: tl.constexpr, +): + last_cumsum = 0 + off_cnt = num_experts * num_experts + for i in range(1, num_experts + 1): + token_cnt = tl.load(tokens_cnts_ptr + off_cnt + i - 1) + last_cumsum = last_cumsum + tl.cdiv(token_cnt, block_size) * block_size + tl.store(cumsum_ptr + i, last_cumsum) + tl.store(total_tokens_post_pad_ptr, last_cumsum) + + +@triton.jit +def moe_align_block_size_stage4( + topk_ids_ptr, + sorted_token_ids_ptr, + expert_ids_ptr, + tokens_cnts_ptr, + cumsum_ptr, + num_experts: tl.constexpr, + block_size: tl.constexpr, + numel: tl.constexpr, + tokens_per_thread: tl.constexpr, +): + pid = tl.program_id(0) + start_idx = tl.load(cumsum_ptr + pid) + end_idx = tl.load(cumsum_ptr + pid + 1) + + for i in range(start_idx, end_idx, block_size): + tl.store(expert_ids_ptr + i // block_size, pid) + + start_idx = pid * tokens_per_thread + off_t = pid * num_experts + + for i in range(start_idx, tl.minimum(start_idx + tokens_per_thread, numel)): + expert_id = tl.load(topk_ids_ptr + i) + token_cnt = tl.load(tokens_cnts_ptr + off_t + expert_id) + rank_post_pad = token_cnt + tl.load(cumsum_ptr + expert_id) + tl.store(sorted_token_ids_ptr + rank_post_pad, i) + tl.store(tokens_cnts_ptr + off_t + expert_id, token_cnt + 1) + + +# Triton implementation based on: +# https://github.com/sgl-project/sglang/commit/ba5112ff691d791a9e38c6c71f59324a5fcb49d0 +def moe_align_block_size_triton( + topk_ids: torch.Tensor, + num_experts: int, + block_size: int, + sorted_token_ids: torch.Tensor, + expert_ids: torch.Tensor, + num_tokens_post_pad: torch.Tensor, +) -> None: + numel = topk_ids.numel() + grid = (num_experts,) + tokens_cnts = torch.zeros( + (num_experts + 1, num_experts), dtype=torch.int32, device=topk_ids.device + ) + cumsum = torch.zeros((num_experts + 1,), dtype=torch.int32, device=topk_ids.device) + tokens_per_thread = ceil_div(numel, num_experts) + + moe_align_block_size_stage1[grid]( + topk_ids, + tokens_cnts, + num_experts, + numel, + tokens_per_thread, + ) + moe_align_block_size_stage2[grid]( + tokens_cnts, + num_experts, + ) + moe_align_block_size_stage3[(1,)]( + num_tokens_post_pad, + tokens_cnts, + cumsum, + num_experts, + block_size, + ) + moe_align_block_size_stage4[grid]( + topk_ids, + sorted_token_ids, + expert_ids, + tokens_cnts, + cumsum, + num_experts, + block_size, + numel, + tokens_per_thread, + ) + + +def moe_align_block_size( + topk_ids: torch.Tensor, + block_size: int, + num_experts: int, + expert_map: Optional[torch.Tensor] = None, + pad_sorted_ids: bool = False, +) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + """ + Aligns the token distribution across experts to be compatible with block + size for matrix multiplication. + + Parameters: + - topk_ids: A tensor of shape [total_tokens, top_k] representing the + top-k expert indices for each token. + - block_size: The block size used in block matrix multiplication. + - num_experts: The total number of experts. + - expert_map: A tensor of shape [num_experts] that maps the expert index + from the global space to the local index space of the current + expert parallel shard. If the expert is not in the current expert + parallel shard, the mapping is set to -1. + - pad_sorted_ids: A flag indicating whether the sorted_token_ids length + should be padded to a multiple of block_size, + + Returns: + - sorted_token_ids: A tensor containing the sorted token indices according + to their allocated expert. + - expert_ids: A tensor indicating the assigned expert index for each block. + - num_tokens_post_padded: The total number of tokens after padding, + ensuring divisibility by block_size. + + This function pads the number of tokens that each expert needs to process + so that it is divisible by block_size. + Padding ensures that during block matrix multiplication, the dimensions + align correctly. + + Example: + Given topk_ids = [[2, 3, 4], [1, 2, 4], [1, 3, 4], [1, 2, 3]], + block_size = 4, and num_experts = 4: + - We initially have 12 tokens (after repeating 'top_k' times) and 4 experts, + with each expert needing to process 3 tokens. + - As block_size is 4, we pad 1 token for each expert. + - First, flatten topk_ids to [2, 3, 4, 1, 2, 4, 1, 3, 4, 1, 2, 3]. + - Then append padding tokens [12, 12, 12, 12] for each block. + - After sorting by expert index, we obtain token_ids + [3, 6, 9, 12, 0, 4, 10, 12, 1, 7, 11, 12, 2, 5, 8, 12]. + Tokens 12 are non-existent (padding) and are ignored in + the subsequent matrix multiplication. + - The padding ensures that the total number of tokens is now divisible + by block_size for proper block matrix operations. + """ + max_num_tokens_padded = topk_ids.numel() + num_experts * (block_size - 1) + if pad_sorted_ids: + max_num_tokens_padded = round_up(max_num_tokens_padded, block_size) + sorted_ids = torch.empty( + (max_num_tokens_padded,), dtype=torch.int32, device=topk_ids.device + ) + sorted_ids.fill_(topk_ids.numel()) + max_num_m_blocks = triton.cdiv(max_num_tokens_padded, block_size) + # Expert ids must be zeroed out to prevent index out of bounds error while + # mapping global expert ids to local expert ids in expert parallelism. + expert_ids = torch.zeros( + (max_num_m_blocks,), dtype=torch.int32, device=topk_ids.device + ) + num_tokens_post_pad = torch.empty((1), dtype=torch.int32, device=topk_ids.device) + if num_experts >= 224: + if VLLM_ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON or num_experts != 256: + moe_align_block_size_triton( + topk_ids, + num_experts, + block_size, + sorted_ids, + expert_ids, + num_tokens_post_pad, + ) + else: + # Currently requires num_experts=256 + ops.sgl_moe_align_block_size( + topk_ids, + num_experts, + block_size, + sorted_ids, + expert_ids, + num_tokens_post_pad, + ) + else: + ops.moe_align_block_size( + topk_ids, + num_experts, + block_size, + sorted_ids, + expert_ids, + num_tokens_post_pad, + ) + if expert_map is not None: + expert_ids = expert_map[expert_ids] + + return sorted_ids, expert_ids, num_tokens_post_pad diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/platforms.py b/build/torch26-cxx98-cu126-aarch64-linux/moe/platforms.py new file mode 100644 index 0000000000000000000000000000000000000000..db83d0f2b0fadb15c14e8887ac27e4a07b6c4d62 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/platforms.py @@ -0,0 +1,99 @@ +import random +from abc import ABC, abstractmethod +from functools import lru_cache +from typing import NamedTuple, Optional + +import torch + +IS_ROCM = torch.version.hip is not None + + +class DeviceCapability(NamedTuple): + major: int + minor: int + + def as_version_str(self) -> str: + return f"{self.major}.{self.minor}" + + def to_int(self) -> int: + """ + Express device capability as an integer ````. + + It is assumed that the minor version is always a single digit. + """ + assert 0 <= self.minor < 10 + return self.major * 10 + self.minor + + +class Platform(ABC): + simple_compile_backend: str = "inductor" + + @classmethod + @abstractmethod + def get_device_name(cls, device_id: int = 0) -> str: ... + + @abstractmethod + def is_cuda_alike(self) -> bool: + """Stateless version of :func:`torch.cuda.is_available`.""" + ... + + @abstractmethod + def is_rocm(self): ... + + @classmethod + def seed_everything(cls, seed: Optional[int] = None) -> None: + """ + Set the seed of each random module. + `torch.manual_seed` will set seed on all devices. + + Loosely based on: https://github.com/Lightning-AI/pytorch-lightning/blob/2.4.0/src/lightning/fabric/utilities/seed.py#L20 + """ + if seed is not None: + random.seed(seed) + torch.manual_seed(seed) + try: + import numpy as np + + np.random.seed(seed) + except ImportError: + pass + + +class CudaPlatform(Platform): + @classmethod + def get_device_capability(cls, device_id: int = 0) -> DeviceCapability: + major, minor = torch.cuda.get_device_capability(device_id) + return DeviceCapability(major=major, minor=minor) + + @classmethod + @lru_cache(maxsize=8) + def get_device_name(cls, device_id: int = 0) -> str: + return torch.cuda.get_device_name(0) + + def is_cuda_alike(self) -> bool: + return True + + def is_rocm(self): + return False + + +class RocmPlatform(Platform): + @classmethod + @lru_cache(maxsize=8) + def get_device_capability(cls, device_id: int = 0) -> DeviceCapability: + major, minor = torch.cuda.get_device_capability(device_id) + return DeviceCapability(major=major, minor=minor) + + @classmethod + @lru_cache(maxsize=8) + def get_device_name(cls, device_id: int = 0) -> str: + return torch.cuda.get_device_name(device_id) + + def is_cuda_alike(self) -> bool: + return True + + def is_rocm(self): + return True + + +current_platform = RocmPlatform() if IS_ROCM else CudaPlatform() diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/scalar_type.py b/build/torch26-cxx98-cu126-aarch64-linux/moe/scalar_type.py new file mode 100644 index 0000000000000000000000000000000000000000..ea749fe8247b6846620ccbba30ddf48d914ca4e1 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/scalar_type.py @@ -0,0 +1,336 @@ +import functools +import struct +from dataclasses import dataclass +from enum import Enum +from typing import Optional, Union + + +# Mirrors enum in `core/scalar_type.hpp` +class NanRepr(Enum): + NONE = 0 # nans are not supported + IEEE_754 = 1 # nans are: Exp all 1s, mantissa not all 0s + EXTD_RANGE_MAX_MIN = 2 # nans are: Exp all 1s, mantissa all 1s + + +# This ScalarType class is a parallel implementation of the C++ ScalarType +# class found in csrc/core/scalar_type.hpp. These two classes should be kept +# in sync until the inductor fully supports custom C++ classes. +@dataclass(frozen=True) +class ScalarType: + """ + ScalarType can represent a wide range of floating point and integer + types, in particular it can be used to represent sub-byte data types + (something that torch.dtype currently does not support). It is also + capable of representing types with a bias, i.e.: + `stored_value = value + bias`, + this is useful for quantized types (e.g. standard GPTQ 4bit uses a bias + of 8). The implementation for this class can be found in + csrc/core/scalar_type.hpp, these type signatures should be kept in sync + with that file. + """ + + exponent: int + """ + Number of bits in the exponent if this is a floating point type + (zero if this an integer type) + """ + + mantissa: int + """ + Number of bits in the mantissa if this is a floating point type, + or the number bits representing an integer excluding the sign bit if + this an integer type. + """ + + signed: bool + "If the type is signed (i.e. has a sign bit)" + + bias: int + """ + bias used to encode the values in this scalar type + (value = stored_value - bias, default 0) for example if we store the + type as an unsigned integer with a bias of 128 then the value 0 will be + stored as 128 and -1 will be stored as 127 and 1 will be stored as 129. + """ + + _finite_values_only: bool = False + """ + Private: if infs are supported, used `has_infs()` instead. + """ + + nan_repr: NanRepr = NanRepr.IEEE_754 + """ + How NaNs are represent in this scalar type, returns NanRepr value. + (not applicable for integer types) + """ + + def _floating_point_max_int(self) -> int: + assert ( + self.mantissa <= 52 and self.exponent <= 11 + ), f"Cannot represent max/min as a double for type {self.__str__()}" + + max_mantissa = (1 << self.mantissa) - 1 + if self.nan_repr == NanRepr.EXTD_RANGE_MAX_MIN: + max_mantissa = max_mantissa - 1 + + max_exponent = (1 << self.exponent) - 2 + if self.nan_repr == NanRepr.EXTD_RANGE_MAX_MIN or self.nan_repr == NanRepr.NONE: + assert ( + self.exponent < 11 + ), f"Cannot represent max/min as a double for type {self.__str__()}" + max_exponent = max_exponent + 1 + + # adjust the exponent to match that of a double + # for now we assume the exponent bias is the standard 2^(e-1) -1, (where + # e is the exponent bits), there is some precedent for non-standard + # biases, example `float8_e4m3b11fnuz` here: + # https://github.com/jax-ml/ml_dtypes but to avoid premature over + # complication we are just assuming the standard exponent bias until + # there is a need to support non-standard biases + exponent_bias = (1 << (self.exponent - 1)) - 1 + exponent_bias_double = (1 << 10) - 1 # double e = 11 + + max_exponent_double = max_exponent - exponent_bias + exponent_bias_double + + # shift the mantissa and exponent into the proper positions for an + # IEEE double and bitwise-or them together. + return (max_mantissa << (52 - self.mantissa)) | (max_exponent_double << 52) + + def _floating_point_max(self) -> float: + double_raw = self._floating_point_max_int() + return struct.unpack("!d", struct.pack("!Q", double_raw))[0] + + def _raw_max(self) -> Union[int, float]: + if self.is_floating_point(): + return self._floating_point_max() + else: + assert ( + self.size_bits < 64 or self.size_bits == 64 and self.is_signed() + ), "Cannot represent max as an int" + return (1 << self.mantissa) - 1 + + def _raw_min(self) -> Union[int, float]: + if self.is_floating_point(): + assert ( + self.is_signed() + ), "We currently assume all floating point types are signed" + sign_bit_double = 1 << 63 + + max_raw = self._floating_point_max_int() + min_raw = max_raw | sign_bit_double + return struct.unpack("!d", struct.pack("!Q", min_raw))[0] + else: + assert ( + not self.is_signed() or self.size_bits <= 64 + ), "Cannot represent min as a int64_t" + + if self.is_signed(): + return -(1 << (self.size_bits - 1)) + else: + return 0 + + @functools.cached_property + def id(self) -> int: + """ + Convert the ScalarType to an int which can be passed to pytorch custom + ops. This layout of the int must be kept in sync with the C++ + ScalarType's from_id method. + """ + val = 0 + offset = 0 + + def or_and_advance(member, bit_width): + nonlocal val + nonlocal offset + bit_mask = (1 << bit_width) - 1 + val = val | (int(member) & bit_mask) << offset + offset = offset + bit_width + + or_and_advance(self.exponent, 8) + or_and_advance(self.mantissa, 8) + or_and_advance(self.signed, 1) + or_and_advance(self.bias, 32) + or_and_advance(self._finite_values_only, 1) + or_and_advance(self.nan_repr.value, 8) + + assert offset <= 64, f"ScalarType fields too big {offset} to fit into an int64" + + return val + + @property + def size_bits(self) -> int: + return self.exponent + self.mantissa + int(self.signed) + + def min(self) -> Union[int, float]: + """ + Min representable value for this scalar type. + (accounting for bias if there is one) + """ + return self._raw_min() - self.bias + + def max(self) -> Union[int, float]: + """ + Max representable value for this scalar type. + (accounting for bias if there is one) + """ + return self._raw_max() - self.bias + + def is_signed(self) -> bool: + """ + If the type is signed (i.e. has a sign bit), same as `signed` + added for consistency with: + https://pytorch.org/docs/stable/generated/torch.Tensor.is_signed.html + """ + return self.signed + + def is_floating_point(self) -> bool: + "If the type is a floating point type" + return self.exponent != 0 + + def is_integer(self) -> bool: + "If the type is an integer type" + return self.exponent == 0 + + def has_bias(self) -> bool: + "If the type has a non-zero bias" + return self.bias != 0 + + def has_infs(self) -> bool: + "If the type is floating point and supports infinity" + return not self._finite_values_only + + def has_nans(self) -> bool: + return self.nan_repr != NanRepr.NONE.value + + def is_ieee_754(self) -> bool: + """ + If the type is a floating point type that follows IEEE 754 + conventions + """ + return self.nan_repr == NanRepr.IEEE_754.value and not self._finite_values_only + + def __str__(self) -> str: + """ + naming generally follows: https://github.com/jax-ml/ml_dtypes + for floating point types (leading f) the scheme is: + `float_em[flags]` + flags: + - no-flags: means it follows IEEE 754 conventions + - f: means finite values only (no infinities) + - n: means nans are supported (non-standard encoding) + for integer types the scheme is: + `[u]int[b]` + - if bias is not present it means its zero + """ + if self.is_floating_point(): + ret = ( + "float" + + str(self.size_bits) + + "_e" + + str(self.exponent) + + "m" + + str(self.mantissa) + ) + + if not self.is_ieee_754(): + if self._finite_values_only: + ret = ret + "f" + if self.nan_repr != NanRepr.NONE: + ret = ret + "n" + + return ret + else: + ret = ("int" if self.is_signed() else "uint") + str(self.size_bits) + if self.has_bias(): + ret = ret + "b" + str(self.bias) + return ret + + def __repr__(self) -> str: + return "ScalarType." + self.__str__() + + # __len__ needs to be defined (and has to throw TypeError) for pytorch's + # opcheck to work. + def __len__(self) -> int: + raise TypeError + + # + # Convenience Constructors + # + + @classmethod + def int_(cls, size_bits: int, bias: Optional[int]) -> "ScalarType": + "Create a signed integer scalar type (size_bits includes sign-bit)." + ret = cls(0, size_bits - 1, True, bias if bias else 0) + ret.id # noqa B018: make sure the id is cached + return ret + + @classmethod + def uint(cls, size_bits: int, bias: Optional[int]) -> "ScalarType": + """Create a unsigned integer scalar type.""" + ret = cls(0, size_bits, False, bias if bias else 0) + ret.id # noqa B018: make sure the id is cached + return ret + + @classmethod + def float_IEEE754(cls, exponent: int, mantissa: int) -> "ScalarType": + """ + Create a standard floating point type + (i.e. follows IEEE 754 conventions). + """ + assert mantissa > 0 and exponent > 0 + ret = cls(exponent, mantissa, True, 0) + ret.id # noqa B018: make sure the id is cached + return ret + + @classmethod + def float_( + cls, exponent: int, mantissa: int, finite_values_only: bool, nan_repr: NanRepr + ) -> "ScalarType": + """ + Create a non-standard floating point type + (i.e. does not follow IEEE 754 conventions). + """ + assert mantissa > 0 and exponent > 0 + assert nan_repr != NanRepr.IEEE_754, ( + "use `float_IEEE754` constructor for floating point types that " + "follow IEEE 754 conventions" + ) + ret = cls(exponent, mantissa, True, 0, finite_values_only, nan_repr) + ret.id # noqa B018: make sure the id is cached + return ret + + +# naming generally follows: https://github.com/jax-ml/ml_dtypes +# for floating point types (leading f) the scheme is: +# `float_em[flags]` +# flags: +# - no-flags: means it follows IEEE 754 conventions +# - f: means finite values only (no infinities) +# - n: means nans are supported (non-standard encoding) +# for integer types the scheme is: +# `[u]int[b]` +# - if bias is not present it means its zero + + +class scalar_types: + int4 = ScalarType.int_(4, None) + uint4 = ScalarType.uint(4, None) + int8 = ScalarType.int_(8, None) + uint8 = ScalarType.uint(8, None) + float8_e4m3fn = ScalarType.float_(4, 3, True, NanRepr.EXTD_RANGE_MAX_MIN) + float8_e5m2 = ScalarType.float_IEEE754(5, 2) + float16_e8m7 = ScalarType.float_IEEE754(8, 7) + float16_e5m10 = ScalarType.float_IEEE754(5, 10) + + # fp6, https://github.com/usyd-fsalab/fp6_llm/tree/main + float6_e3m2f = ScalarType.float_(3, 2, True, NanRepr.NONE) + + # "gptq" types + uint2b2 = ScalarType.uint(2, 2) + uint3b4 = ScalarType.uint(3, 4) + uint4b8 = ScalarType.uint(4, 8) + uint8b128 = ScalarType.uint(8, 128) + + # colloquial names + bfloat16 = float16_e8m7 + float16 = float16_e5m10 diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/utils/__init__.py b/build/torch26-cxx98-cu126-aarch64-linux/moe/utils/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/utils/custom_ops.py b/build/torch26-cxx98-cu126-aarch64-linux/moe/utils/custom_ops.py new file mode 100644 index 0000000000000000000000000000000000000000..26ef7c23d1bea7a162d81746ac041baefb472f46 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/utils/custom_ops.py @@ -0,0 +1,68 @@ +from typing import Callable, Optional + +import torch +from torch.library import Library + +from .._ops import ops +from ..platforms import current_platform + + +# Some backends use pytorch version < 2.4.0 which doesn't +# support `torch.library.custom_op`. +def supports_custom_op() -> bool: + return hasattr(torch.library, "custom_op") + + +# TODO: we probably need a shorthand for the ops name. +vllm_lib = Library(ops.__name__.split(".")[-1], "FRAGMENT") + + +def direct_register_custom_op( + op_name: str, + op_func: Callable, + mutates_args: list[str], + fake_impl: Optional[Callable] = None, + target_lib: Optional[Library] = None, + dispatch_key: str = "CUDA", +): + """ + `torch.library.custom_op` can have significant overhead because it + needs to consider complicated dispatching logic. This function + directly registers a custom op and dispatches it to the CUDA backend. + See https://gist.github.com/youkaichao/ecbea9ec9fc79a45d2adce1784d7a9a5 + for more details. + + By default, the custom op is registered to the vLLM library. If you + want to register it to a different library, you can pass the library + object to the `target_lib` argument. + + IMPORTANT: the lifetime of the operator is tied to the lifetime of the + library object. If you want to bind the operator to a different library, + make sure the library object is alive when the operator is used. + """ + if not supports_custom_op(): + from ..platforms import current_platform + + assert not current_platform.is_cuda_alike(), ( + "cuda platform needs torch>=2.4 to support custom op, " + "chances are you are using an old version of pytorch " + "or a custom build of pytorch. It is recommended to " + "use vLLM in a fresh new environment and let it install " + "the required dependencies." + ) + return + + import torch.library + + if hasattr(torch.library, "infer_schema"): + schema_str = torch.library.infer_schema(op_func, mutates_args=mutates_args) + else: + # for pytorch 2.4 + import torch._custom_op.impl + + schema_str = torch._custom_op.impl.infer_schema(op_func, mutates_args) + my_lib = target_lib or vllm_lib + my_lib.define(op_name + schema_str) + my_lib.impl(op_name, op_func, dispatch_key=dispatch_key) + if fake_impl is not None: + my_lib._register_fake(op_name, fake_impl) diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/utils/marlin_utils.py b/build/torch26-cxx98-cu126-aarch64-linux/moe/utils/marlin_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..f27b71af67a6f5f18c8dba271926e3d40d518879 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/utils/marlin_utils.py @@ -0,0 +1,306 @@ +from typing import List, Optional, Tuple + +import numpy +import torch + +from ..scalar_type import ScalarType, scalar_types +from .quant_utils import pack_cols, unpack_cols + +GPTQ_MARLIN_TILE = 16 +GPTQ_MARLIN_MIN_THREAD_N = 64 +GPTQ_MARLIN_MIN_THREAD_K = 128 +GPTQ_MARLIN_MAX_PARALLEL = 16 + +GPTQ_MARLIN_24_TILE = 16 +GPTQ_MARLIN_24_MIN_THREAD_N = 128 +GPTQ_MARLIN_24_MIN_THREAD_K = 128 +GPTQ_MARLIN_24_MAX_PARALLEL = 64 + +GPTQ_MARLIN_24_SUPPORTED_QUANT_TYPES = [scalar_types.uint4b8, scalar_types.uint8b128] +GPTQ_MARLIN_24_SUPPORTED_GROUP_SIZES = [-1, 128] + +MARLIN_QQQ_TILE = 16 +MARLIN_QQQ_MIN_THREAD_N = 64 +MARLIN_QQQ_MIN_THREAD_K = 128 +MARLIN_QQQ_MAX_PARALLEL = 16 + +MARLIN_QQQ_SUPPORTED_NUM_BITS = [4] +MARLIN_QQQ_SUPPORTED_GROUP_SIZES = [-1, 128] +MARLIN_QQQ_SUPPORTED_SYM = [True] + +MARLIN_SUPPORTED_GROUP_SIZES = [-1, 32, 64, 128] + +# In case there is a performance issue with Marlin, the variable below can be +# changed to False, which allows Marlin to perform global reductions in fp16 +# precision (instead of fp32), and therefore, save on some memory movements. +USE_FP32_REDUCE_DEFAULT = True + + +# For binary size and compile time, we don't support the same types for with and +# without runtime zero-point. We support common cases, i.e. AWQ and GPTQ. +# TODO: we may want to move this into the C++ so its closer to the actual impl +def query_marlin_supported_quant_types( + has_zp: bool, device_capability: Optional[int] = None +): + if device_capability is None: + capability_tuple = torch.cuda.get_device_capability() + device_capability = capability_tuple[0] * 10 + capability_tuple[1] + + if device_capability < 80: + return [] + + if has_zp: + # AWQ style, unsigned + runtime zero-point + return [scalar_types.uint4, scalar_types.uint8] + else: + # GPTQ style, unsigned + symmetric bias + # TODO: once fp8_marlin is merged into "gptq_marlin" we should be able + # to add `scalar_types.float8_e4m3fn` here + return [scalar_types.uint4b8, scalar_types.uint8b128] + + +def _check_marlin_supported( + quant_type: ScalarType, + group_size: Optional[int], + has_zp: bool, + device_capability: Optional[int] = None, +) -> Tuple[bool, Optional[str]]: + + if device_capability is None: + capability_tuple = torch.cuda.get_device_capability() + device_capability = capability_tuple[0] * 10 + capability_tuple[1] + + supported_types = query_marlin_supported_quant_types(has_zp, device_capability) + + if quant_type not in supported_types: + return ( + False, + f"Marlin does not support weight_bits = {quant_type}. " + f"Only types = {supported_types} " + f"are supported (for group_size = {group_size}, " + f"device_capability = {device_capability}, zp = {has_zp}).", + ) + if group_size is None or group_size not in MARLIN_SUPPORTED_GROUP_SIZES: + return ( + False, + f"Marlin does not support group_size = {group_size}. " + f"Only group_sizes = {MARLIN_SUPPORTED_GROUP_SIZES} " + "are supported.", + ) + + return True, None + + +def check_marlin_supported( + quant_type: ScalarType, + group_size: int, + has_zp: bool = False, + device_capability: Optional[int] = None, +) -> bool: + cond, _ = _check_marlin_supported(quant_type, group_size, has_zp, device_capability) + return cond + + +def verify_marlin_supported( + quant_type: ScalarType, group_size: int, has_zp: bool = False +) -> None: + cond, err_msg = _check_marlin_supported(quant_type, group_size, has_zp) + if not cond: + assert err_msg is not None + raise ValueError(err_msg) + + +def verify_marlin_supports_shape( + output_size_per_partition: int, + input_size_per_partition: int, + input_size: int, + group_size: int, +) -> None: + + # Validate output_size_per_partition + if output_size_per_partition % GPTQ_MARLIN_MIN_THREAD_N != 0: + raise ValueError( + f"Weight output_size_per_partition = " + f"{output_size_per_partition} is not divisible by " + f" min_thread_n = {GPTQ_MARLIN_MIN_THREAD_N}. " + "Consider reducing tensor_parallel_size or running " + "with --quantization gptq." + ) + + # Validate input_size_per_partition + if input_size_per_partition % GPTQ_MARLIN_MIN_THREAD_K != 0: + raise ValueError( + f"Weight input_size_per_partition = " + f"{input_size_per_partition} is not divisible " + f"by min_thread_k = {GPTQ_MARLIN_MIN_THREAD_K}. " + "Consider reducing tensor_parallel_size or running " + "with --quantization gptq." + ) + + if group_size < input_size and input_size_per_partition % group_size != 0: + raise ValueError( + f"Weight input_size_per_partition = {input_size_per_partition}" + f" is not divisible by group_size = {group_size}." + "Consider reducing tensor_parallel_size or running " + "with --quantization gptq." + ) + + +def check_marlin_supports_shape( + output_size_per_partition: int, + input_size_per_partition: int, + input_size: int, + group_size: int, +) -> Tuple[bool, Optional[str]]: + try: + verify_marlin_supports_shape( + output_size_per_partition, input_size_per_partition, input_size, group_size + ) + except ValueError as e: + return False, e.__str__() + return True, None + + +def marlin_make_workspace( + output_size_per_partition: int, device: torch.device +) -> torch.Tensor: + max_workspace_size = ( + output_size_per_partition // GPTQ_MARLIN_MIN_THREAD_N + ) * GPTQ_MARLIN_MAX_PARALLEL + + return torch.zeros( + max_workspace_size, dtype=torch.int, device=device, requires_grad=False + ) + + +def marlin_is_k_full(act_order: bool, is_row_parallel: bool) -> bool: + return (not act_order) or (act_order and not is_row_parallel) + + +def marlin_repeat_scales_on_all_ranks( + act_order: bool, group_size: int, is_row_parallel: bool +) -> bool: + # Need to repeat scales on every rank if act_ordering or + # channelwise and RowParallelLinear + is_channelwise = group_size == -1 + return act_order or (is_channelwise and is_row_parallel) + + +def marlin_make_empty_g_idx(device: torch.device) -> torch.Tensor: + return torch.nn.Parameter( + torch.empty(0, dtype=torch.int, device=device), requires_grad=False + ) + + +def marlin_make_empty_zp(device: torch.device) -> torch.Tensor: + return torch.nn.Parameter( + torch.empty(0, dtype=torch.int, device=device), requires_grad=False + ) + + +def marlin_sort_g_idx(g_idx: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]: + g_idx_sort_indices = torch.argsort(g_idx).to(torch.int) + return g_idx[g_idx_sort_indices], g_idx_sort_indices + + +def get_scale_perms(): + scale_perm: List[int] = [] + for i in range(8): + scale_perm.extend([i + 8 * j for j in range(8)]) + scale_perm_single: List[int] = [] + for i in range(4): + scale_perm_single.extend([2 * i + j for j in [0, 1, 8, 9, 16, 17, 24, 25]]) + return scale_perm, scale_perm_single + + +def marlin_permute_scales( + s: torch.Tensor, size_k: int, size_n: int, group_size: int +) -> torch.Tensor: + + scale_perm, scale_perm_single = get_scale_perms() + if group_size < size_k and group_size != -1: + s = s.reshape((-1, len(scale_perm)))[:, scale_perm] + else: + s = s.reshape((-1, len(scale_perm_single)))[:, scale_perm_single] + s = s.reshape((-1, size_n)).contiguous() + + return s + + +def marlin_moe_permute_scales( + s: torch.Tensor, + size_k: int, + size_n: int, + group_size: int, +): + num_experts = s.shape[0] + output = torch.empty( + (num_experts, s.shape[1], s.shape[2]), + device=s.device, + dtype=s.dtype, + ) + + for e in range(num_experts): + output[e] = marlin_permute_scales(s[e], size_k, size_n, group_size) + return output + + +def marlin_zero_points( + zp: torch.Tensor, size_k: int, size_n: int, num_bits: int +) -> torch.Tensor: + # Permute zero-points in a similar way to scales, but do not use the + # "single" permutation, since zero-points are applied on every MMA + scale_perm, _ = get_scale_perms() + zp = zp.reshape((-1, len(scale_perm)))[:, scale_perm] + + # Interleave column dim (for the dequantize code) and pack it to int32 + if num_bits == 4: + interleave = numpy.array([0, 2, 4, 6, 1, 3, 5, 7]) + elif num_bits == 8: + interleave = numpy.array([0, 2, 1, 3]) + else: + raise Exception("num_bits must be 4 or 8, got {}".format(num_bits)) + + zp = zp.reshape((-1, len(interleave)))[:, interleave].ravel() + zp = zp.reshape((-1, size_n)).contiguous() + zp = pack_cols(zp, num_bits, size_k, size_n) + + return zp + + +def awq_to_marlin_zero_points( + q_zp_packed: torch.Tensor, size_k: int, size_n: int, num_bits: int +) -> torch.Tensor: + # AWQ zero-points are quantized and packed on the column dim. + # In addition, the values are permuted based on dequantizer. + # Here we undo both of these, and then apply marlin permutation + # and pack it back. + q_zp = unpack_cols(q_zp_packed, num_bits, size_k, size_n) + + # Undo interleaving (use argsort(..) to get inverse perm) + if num_bits == 4: + undo_interleave = numpy.argsort(numpy.array([0, 2, 4, 6, 1, 3, 5, 7])) + elif num_bits == 8: + undo_interleave = numpy.argsort(numpy.array([0, 2, 1, 3])) + else: + raise Exception("num_bits must be 4 or 8, got {}".format(num_bits)) + + q_zp = q_zp.reshape((-1, len(undo_interleave)))[:, undo_interleave].ravel() + q_zp = q_zp.reshape((-1, size_n)).contiguous() + + marlin_zp = marlin_zero_points(q_zp, size_k, size_n, num_bits) + return marlin_zp + + +def moe_awq_to_marlin_zero_points( + q_zp_packed: torch.Tensor, size_k: int, size_n: int, num_bits: int +): + num_experts = q_zp_packed.shape[0] + output = torch.empty( + (num_experts, q_zp_packed.shape[1], q_zp_packed.shape[2]), + device=q_zp_packed.device, + dtype=q_zp_packed.dtype, + ) + for e in range(num_experts): + output[e] = awq_to_marlin_zero_points(q_zp_packed[e], size_k, size_n, num_bits) + return output diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/utils/marlin_utils_test.py b/build/torch26-cxx98-cu126-aarch64-linux/moe/utils/marlin_utils_test.py new file mode 100644 index 0000000000000000000000000000000000000000..b03555f5bffa8d48187d37e53f17e6e3138472cc --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/utils/marlin_utils_test.py @@ -0,0 +1,161 @@ +"""Utility functions used for tests and benchmarks""" + +from typing import List, Optional + +import numpy as np +import torch + +from ..scalar_type import ScalarType +from .marlin_utils import GPTQ_MARLIN_TILE, marlin_permute_scales, marlin_zero_points +from .quant_utils import ( + get_pack_factor, + gptq_quantize_weights, + quantize_weights, + sort_weights, +) + + +class MarlinWorkspace: + + def __init__(self, out_features, min_thread_n, max_parallel): + assert ( + out_features % min_thread_n == 0 + ), "out_features = {} is undivisible by min_thread_n = {}".format( + out_features, min_thread_n + ) + + max_workspace_size = (out_features // min_thread_n) * max_parallel + + self.scratch = torch.zeros(max_workspace_size, dtype=torch.int, device="cuda") + + +def marlin_permute_weights(q_w, size_k, size_n, perm, tile=GPTQ_MARLIN_TILE): + assert q_w.shape == (size_k, size_n) + assert size_k % tile == 0, f"size_k = {size_k}, tile = {tile}" + assert size_n % tile == 0, f"size_k = {size_n}, tile = {tile}" + + # Permute weights to 16x64 marlin tiles + q_w = q_w.reshape((size_k // tile, tile, size_n // tile, tile)) + q_w = q_w.permute((0, 2, 1, 3)) + q_w = q_w.reshape((size_k // tile, size_n * tile)) + + q_w = q_w.reshape((-1, perm.numel()))[:, perm].reshape(q_w.shape) + + return q_w + + +def marlin_weights(q_w, size_k, size_n, num_bits, perm): + # Permute + q_w = marlin_permute_weights(q_w, size_k, size_n, perm) + + # Pack + pack_factor = get_pack_factor(num_bits) + orig_device = q_w.device + + q_w = q_w.cpu().numpy().astype(np.uint32) + + q_packed = np.zeros((q_w.shape[0], q_w.shape[1] // pack_factor), dtype=np.uint32) + for i in range(pack_factor): + q_packed |= q_w[:, i::pack_factor] << num_bits * i + + q_packed = torch.from_numpy(q_packed.astype(np.int32)).to(orig_device) + + return q_packed + + +def get_weight_perm(num_bits: int): + perm_list: List[int] = [] + for i in range(32): + perm1: List[int] = [] + col = i // 4 + for block in [0, 1]: + for row in [ + 2 * (i % 4), + 2 * (i % 4) + 1, + 2 * (i % 4 + 4), + 2 * (i % 4 + 4) + 1, + ]: + perm1.append(16 * row + col + 8 * block) + for j in range(4): + perm_list.extend([p + 256 * j for p in perm1]) + + perm = np.array(perm_list) + + if num_bits == 4: + interleave = np.array([0, 2, 4, 6, 1, 3, 5, 7]) + elif num_bits == 8: + interleave = np.array([0, 2, 1, 3]) + else: + raise Exception("num_bits must be 4 or 8, got {}".format(num_bits)) + + perm = perm.reshape((-1, len(interleave)))[:, interleave].ravel() + perm = torch.from_numpy(perm) + return perm + + +def marlin_quantize( + w: torch.Tensor, + quant_type: ScalarType, + group_size: int, + act_order: bool, + test_perm: Optional[torch.Tensor] = None, +): + size_k, size_n = w.shape + num_bits = quant_type.size_bits + + # Normalize group_size + if group_size == -1: + group_size = size_k + assert group_size <= size_k + + # Quantize (and apply act_order if provided) + w_ref, q_w, s, g_idx, rand_perm = gptq_quantize_weights( + w, quant_type, group_size, act_order, test_perm + ) + + # For act_order, sort the "weights" and "g_idx" so that group ids are + # increasing + sort_indices = torch.empty(0, dtype=torch.int, device=w.device) + if act_order: + q_w, g_idx, sort_indices = sort_weights(q_w, g_idx) + + # Reformat to marlin + weight_perm = get_weight_perm(num_bits) + marlin_q_w = marlin_weights(q_w, size_k, size_n, num_bits, weight_perm) + marlin_s = marlin_permute_scales(s, size_k, size_n, group_size) + + # Create result + res_list = [w_ref, marlin_q_w, marlin_s, g_idx, sort_indices, rand_perm] + for i in range(len(res_list)): + res_list[i] = res_list[i].to(w.device) + + return res_list + + +def awq_marlin_quantize(w: torch.Tensor, quant_type: ScalarType, group_size: int): + size_k, size_n = w.shape + + # Normalize group_size + if group_size == -1: + group_size = size_k + assert group_size <= size_k + + # Detect num groups + assert size_k % group_size == 0 + num_groups = size_k // group_size + + # Quantize with zp + w_ref, q_w, s, zp = quantize_weights(w, quant_type, group_size, zero_points=True) + + # Reformat to marlin + weight_perm = get_weight_perm(quant_type.size_bits) + marlin_q_w = marlin_weights(q_w, size_k, size_n, quant_type.size_bits, weight_perm) + marlin_s = marlin_permute_scales(s, size_k, size_n, group_size) + marlin_zp = marlin_zero_points(zp, num_groups, size_n, quant_type.size_bits) + + # Create result + res_list = [w_ref, marlin_q_w, marlin_s, marlin_zp] + for i in range(len(res_list)): + res_list[i] = res_list[i].to(w.device) + + return res_list diff --git a/build/torch26-cxx98-cu126-aarch64-linux/moe/utils/quant_utils.py b/build/torch26-cxx98-cu126-aarch64-linux/moe/utils/quant_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..65c90821773b60cca2b0f9102ca7ad4e3cd009d5 --- /dev/null +++ b/build/torch26-cxx98-cu126-aarch64-linux/moe/utils/quant_utils.py @@ -0,0 +1,470 @@ +"""This file is used for /tests and /benchmarks""" + +from typing import List, Optional + +import numpy +import torch + +from ..scalar_type import ScalarType, scalar_types + +SUPPORTED_GPTQ_QUANT_TYPES = [scalar_types.uint4b8, scalar_types.uint8b128] +SUPPORTED_GROUP_SIZES = [-1, 32, 64, 128] + +MARLIN_QQQ_SUPPORTED_NUM_BITS = [4] + +# Note: this is a hack. We should update each model to register the +# stacked params and get it from there instead in a future PR. +# fused_name: List[shard_name] +FUSED_LAYER_NAME_MAPPING = { + "qkv_proj": ["q_proj", "k_proj", "v_proj"], + "gate_up_proj": ["gate_proj", "up_proj"], +} + + +def pack_quantized_values_into_int32( + w_q: torch.Tensor, wtype: ScalarType, packed_dim: int = 0 +): + # move dim to pack to the end + perm = (*[i for i in range(len(w_q.shape)) if i != packed_dim], packed_dim) + inv_perm = tuple(perm.index(i) for i in range(len(perm))) + w_q_perm = w_q.permute(perm) + + pack_factor = 32 // wtype.size_bits + mask = (1 << wtype.size_bits) - 1 + + new_shape_perm = list(w_q_perm.shape) + assert w_q_perm.shape[-1] % pack_factor == 0 + new_shape_perm[-1] //= pack_factor + + res = torch.zeros(new_shape_perm, dtype=torch.int32, device=w_q.device) + for i in range(pack_factor): + res |= (w_q_perm[..., i::pack_factor] & mask) << wtype.size_bits * i + + return res.permute(inv_perm) + + +def unpack_quantized_values_into_int32( + w_q: torch.Tensor, wtype: ScalarType, packed_dim: int = 0 +): + # move dim to pack to the end + perm = (*[i for i in range(len(w_q.shape)) if i != packed_dim], packed_dim) + inv_perm = tuple(perm.index(i) for i in range(len(perm))) + w_q_perm = w_q.permute(perm) + + pack_factor = 32 // wtype.size_bits + mask = (1 << wtype.size_bits) - 1 + + new_shape_perm = list(w_q_perm.shape) + new_shape_perm[-1] *= pack_factor + + res = torch.zeros(new_shape_perm, dtype=torch.int32, device=w_q.device) + for i in range(pack_factor): + res[..., i::pack_factor] = (w_q_perm >> wtype.size_bits * i) & mask + + return res.permute(inv_perm) + + +def is_layer_skipped(prefix: str, ignored_layers: List[str]) -> bool: + # prefix: model.layers.0.self_attn.q_proj + # proj_name: q_proj + proj_name = prefix.split(".")[-1] + if proj_name in FUSED_LAYER_NAME_MAPPING: + shard_prefixes = [ + prefix.replace(proj_name, shard_proj_name) + for shard_proj_name in FUSED_LAYER_NAME_MAPPING[proj_name] + ] + + is_skipped = None + for shard_prefix in shard_prefixes: + is_shard_skipped = shard_prefix in ignored_layers + + if is_skipped is None: + is_skipped = is_shard_skipped + elif is_shard_skipped != is_skipped: + raise ValueError( + f"Detected some but not all shards of {prefix} " + "are quantized. All shards of fused layers " + "to have the same precision." + ) + else: + is_skipped = prefix in ignored_layers + + assert is_skipped is not None + return is_skipped + + +def get_pack_factor(num_bits): + assert 32 % num_bits == 0, f"Unsupported num_bits = {num_bits}" + return 32 // num_bits + + +def permute_rows( + q_w: torch.Tensor, + w_ref: torch.Tensor, + group_size: int, + test_perm: Optional[torch.Tensor] = None, +): + assert q_w.shape == w_ref.shape + + orig_device = q_w.device + k_size, _ = q_w.shape + + g_idx = torch.zeros((k_size,), dtype=torch.int32) + for i in range(k_size): + g_idx[i] = i // group_size + + # Simulate act_order by doing a random permutation on K + rand_perm = test_perm if test_perm is not None else torch.randperm(k_size) + + g_idx = g_idx[rand_perm].contiguous() + q_w = q_w[rand_perm, :].contiguous() + w_ref = w_ref[rand_perm, :].contiguous() + + return ( + w_ref.to(device=orig_device), + q_w.to(device=orig_device), + g_idx.to(device=orig_device), + rand_perm.to(device=orig_device), + ) + + +def quantize_weights( + w: torch.Tensor, + quant_type: ScalarType, + group_size: Optional[int], + zero_points: bool = False, + ref_zero_points_after_scales: bool = False, +): + assert ( + quant_type.is_integer() + ), "Floating point quantization may work but has not been tested" + assert not zero_points or group_size is not None, ( + "to have group zero points, group_size must be provided " + "(-1 group_size is channelwise)" + ) + + orig_device = w.device + orig_type = w.dtype + size_k, size_n = w.shape + + assert w.is_floating_point(), "w must be float" + + if group_size == -1: + group_size = size_k + + # Reshape to [groupsize, -1] + if group_size is not None and group_size < size_k: + w = w.reshape((-1, group_size, size_n)) + w = w.permute(1, 0, 2) + w = w.reshape((group_size, -1)) + + # Compute scale for each group + max_val = torch.max(w, 0, keepdim=True).values + min_val = torch.min(w, 0, keepdim=True).values + + max_q_val = quant_type.max() + min_q_val = quant_type.min() + + w_s = torch.Tensor([1.0]).to(w.device) # unscaled case + maybe_w_zp = None + if group_size is not None: + if zero_points: + assert not quant_type.is_signed() and quant_type.max() > 0 + w_s = (max_val - min_val).clamp(min=1e-5) / quant_type.max() + maybe_w_zp = ( + torch.round(torch.abs(min_val / w_s)).clamp(min_q_val, max_q_val).int() + ) + else: + # If the bias is such that there are no possible negative/positive + # values, set the max value to inf to avoid divide by 0 + w_s = torch.max( + abs(max_val / (max_q_val if max_q_val != 0 else torch.inf)), + abs(min_val / (min_q_val if min_q_val != 0 else torch.inf)), + ) + + # Quantize + w_q = torch.round(w / w_s).int() + (maybe_w_zp if zero_points else 0) + w_q = torch.clamp(w_q, min_q_val, max_q_val) + + # Compute ref (dequantized) + # For some kernels (namely Machete) the zero-points are applied after the + # scales are applied, for this case computing the reference in similar way + # allows us to use tighter error tolerances in our unit tests. + if ref_zero_points_after_scales and maybe_w_zp is not None: + w_ref = w_q.to(orig_type) * w_s - maybe_w_zp.to(orig_type) * w_s + else: + w_ref = (w_q - (maybe_w_zp if zero_points else 0)).to(orig_type) * w_s + + if quant_type.has_bias(): + w_q += quant_type.bias + + # Restore original shapes + if group_size is not None and group_size < size_k: + + def reshape_w(w): + w = w.reshape((group_size, -1, size_n)) + w = w.permute(1, 0, 2) + w = w.reshape((size_k, size_n)).contiguous() + return w + + w_q = reshape_w(w_q) + w_ref = reshape_w(w_ref) + w_s = w_s.reshape((-1, size_n)).contiguous() + + if maybe_w_zp is not None: + maybe_w_zp = maybe_w_zp.reshape((-1, size_n)).contiguous() + maybe_w_zp = maybe_w_zp.to(device=orig_device) + + return ( + w_ref.to(device=orig_device), + w_q.to(device=orig_device), + w_s if group_size is not None else None, + maybe_w_zp, + ) + + +def gptq_quantize_weights( + w: torch.Tensor, + quant_type: ScalarType, + group_size: int, + act_order: bool, + test_perm: Optional[torch.Tensor] = None, +): + size_k, _ = w.shape + + assert w.is_floating_point(), "w must be float" + assert ( + quant_type in SUPPORTED_GPTQ_QUANT_TYPES + ), f"Unsupported gptq type = {quant_type}" + assert group_size in SUPPORTED_GROUP_SIZES + [ + size_k + ], f"Unsupported groupsize = {group_size}" + + w_ref, w_q, w_s, _ = quantize_weights(w, quant_type, group_size) + + # Apply act_order + g_idx = torch.empty(0, dtype=torch.int, device=w.device) + rand_perm = torch.empty(0, dtype=torch.int, device=w.device) + if act_order: + assert ( + group_size < size_k + ), "For act_order, groupsize = {} must be less than size_k = {}".format( + group_size, size_k + ) + + w_ref, w_q, g_idx, rand_perm = permute_rows(w_q, w_ref, group_size, test_perm) + + return w_ref, w_q, w_s, g_idx, rand_perm + + +# QQQ employs different quant schemes for per-group and +# per-channel quantization. +def qqq_quantize_weights(w: torch.Tensor, num_bits: int, group_size: int): + orig_device = w.device + size_k, size_n = w.shape + + assert w.is_floating_point(), "w must be float" + assert ( + num_bits in MARLIN_QQQ_SUPPORTED_NUM_BITS + ), f"Unsupported num_bits = {num_bits}" + assert group_size in SUPPORTED_GROUP_SIZES + [ + size_k + ], f"Unsupported groupsize = {group_size}" + + if group_size == -1: + group_size = size_k + assert group_size <= size_k + + if group_size < size_k: + # Reshape to [groupsize, -1] + w = w.reshape((-1, group_size, size_n)) + w = w.permute(1, 0, 2) + w = w.reshape((group_size, -1)) + + max_q_val = 2**num_bits - 1 + half_q_val = (max_q_val + 1) // 2 + + # Compute scale for each group + s_group = torch.max(torch.abs(w), 0, keepdim=True)[0] + s_group *= 2 / max_q_val # 2 => symmetric + + # Quantize + q_w = torch.round(w / s_group).int() + q_w += half_q_val + q_w = torch.clamp(q_w, 0, max_q_val) + # Compute ref (dequantized) + w_ref = (q_w - half_q_val).half() * s_group + + # Restore original shapes + def reshape_w(w): + w = w.reshape((group_size, -1, size_n)) + w = w.permute(1, 0, 2) + w = w.reshape((size_k, size_n)).contiguous() + return w + + q_w = reshape_w(q_w) + w_ref = reshape_w(w_ref) + + # Compute int8 quantization scale for each channel + s_channel = torch.max(torch.abs(w_ref), 0, keepdim=True)[0] + s_channel /= 127.0 + t_int8 = (w_ref / s_channel).round().clamp(-128, 127).to(torch.int8) + w_ref = t_int8.half() * s_channel + s_channel = s_channel.reshape(1, -1).to(dtype=torch.float) + + # Fuse scales + s_group = (s_group.reshape(-1, size_n).contiguous() / s_channel).to( + dtype=torch.half + ) + else: + max_q_val = 2 ** (num_bits - 1) - 1 + + # Compute scale for each channel + s_channel = torch.max(torch.abs(w), 0, keepdim=True)[0] + s_channel /= max_q_val + + # Quantize + q_w = torch.round(w / s_channel).int() + q_w = torch.clamp(q_w, -max_q_val, max_q_val) + # Compute ref (dequantized) + w_ref = q_w.half() * s_channel + + s_group = torch.tensor([], dtype=torch.half) + # div 2 ** (8 - self.bits)) to offset right shift in unpacking + s_channel /= 2 ** (8 - num_bits) + s_channel = s_channel.reshape(-1, size_n).contiguous().to(torch.float) + + return ( + w_ref.to(device=orig_device), + q_w.to(device=orig_device), + s_group.to(device=orig_device), + s_channel.to(device=orig_device), + ) + + +def sort_weights(q_w: torch.Tensor, g_idx: torch.Tensor): + orig_device = q_w.device + + sort_indices = torch.argsort(g_idx).to(dtype=torch.int32) # Sort based on g_idx + + g_idx = g_idx[sort_indices].contiguous() + q_w = q_w[sort_indices, :].contiguous() + + return ( + q_w.to(device=orig_device), + g_idx.to(device=orig_device), + sort_indices.to(device=orig_device), + ) + + +def pack_rows( + q_w: torch.Tensor, + num_bits: int, + size_k: int, + size_n: int, +): + assert q_w.shape == (size_k, size_n) + + pack_factor = get_pack_factor(num_bits) + assert size_k % pack_factor == 0 + + orig_device = q_w.device + + q_w = q_w.cpu().numpy().astype(numpy.uint32) + + q_res = numpy.zeros((size_k // pack_factor, size_n), dtype=numpy.uint32) + + for i in range(pack_factor): + q_res |= q_w[i::pack_factor, :] << num_bits * i + + q_res = torch.from_numpy(q_res.astype(numpy.int32)).to(orig_device) + return q_res + + +def pack_cols( + q_w: torch.Tensor, + num_bits: int, + size_k: int, + size_n: int, +): + assert q_w.shape == (size_k, size_n) + + pack_factor = get_pack_factor(num_bits) + assert size_n % pack_factor == 0 + + orig_device = q_w.device + + q_w = q_w.cpu().numpy().astype(numpy.uint32) + + q_res = numpy.zeros((size_k, size_n // pack_factor), dtype=numpy.uint32) + + for i in range(pack_factor): + q_res |= q_w[:, i::pack_factor] << num_bits * i + + q_res = torch.from_numpy(q_res.astype(numpy.int32)).to(orig_device) + q_res = q_res.contiguous() + + return q_res + + +def unpack_cols( + packed_q_w: torch.Tensor, + num_bits: int, + size_k: int, + size_n: int, +): + pack_factor = get_pack_factor(num_bits) + assert size_n % pack_factor == 0 + assert packed_q_w.shape == ( + size_k, + size_n // pack_factor, + ), "packed_q_w.shape = {} size_k = {}, size_n = {} pack_Factor = {}".format( + packed_q_w.shape, size_k, size_n, pack_factor + ) + + orig_device = packed_q_w.device + + packed_q_w_cpu = packed_q_w.cpu().numpy().astype(numpy.uint32) + q_res = numpy.zeros((size_k, size_n), dtype=numpy.uint32) + + mask = (1 << num_bits) - 1 + for i in range(pack_factor): + vals = packed_q_w_cpu & mask + packed_q_w_cpu >>= num_bits + q_res[:, i::pack_factor] = vals + + q_res = torch.from_numpy(q_res.astype(numpy.int32)).to(orig_device) + q_res = q_res.contiguous() + + return q_res + + +def gptq_pack( + q_w: torch.Tensor, + num_bits: int, + size_k: int, + size_n: int, +): + return pack_rows(q_w, num_bits, size_k, size_n) + + +def awq_pack( + q_w: torch.Tensor, + num_bits: int, + size_k: int, + size_n: int, +): + assert q_w.shape == (size_k, size_n) + + # Interleave column dim (for the dequantize code) and pack it to int32 + if num_bits == 4: + interleave = numpy.array([0, 2, 4, 6, 1, 3, 5, 7]) + elif num_bits == 8: + interleave = numpy.array([0, 2, 1, 3]) + else: + raise Exception("num_bits must be 4 or 8, got {}".format(num_bits)) + + q_w = q_w.reshape((-1, len(interleave)))[:, interleave].ravel() + q_w = q_w.reshape((-1, size_n)).contiguous() + + return pack_cols(q_w, num_bits, size_k, size_n) diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/__init__.py b/build/torch27-cxx11-cu126-aarch64-linux/moe/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..0d9d8227711def6d51c24b5fd779a3d13681d88c --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/__init__.py @@ -0,0 +1,80 @@ +import torch + +from . import layers +from ._ops import ops +from .fp8_utils import per_token_group_quant_fp8, w8a8_block_fp8_matmul +from .fused_marlin_moe import fused_marlin_moe +from .fused_moe import fused_experts, fused_moe, fused_topk, grouped_topk +from .scalar_type import ScalarType, scalar_types + + +def gptq_marlin_moe_repack( + b_q_weight: torch.Tensor, + perm: torch.Tensor, + size_k: int, + size_n: int, + num_bits: int, +) -> torch.Tensor: + num_experts = b_q_weight.shape[0] + assert size_k % 16 == 0 + output = torch.empty( + (num_experts, size_k // 16, size_n * (num_bits // 2)), + device=b_q_weight.device, + dtype=b_q_weight.dtype, + ) + for e in range(num_experts): + output[e] = ops.gptq_marlin_repack( + b_q_weight[e], perm[e], size_k, size_n, num_bits + ) + return output + + +def awq_marlin_moe_repack( + b_q_weight: torch.Tensor, + perm: torch.Tensor, + size_k: int, + size_n: int, + num_bits: int, +) -> torch.Tensor: + num_experts = b_q_weight.shape[0] + assert size_k % 16 == 0 + output = torch.empty( + (num_experts, size_k // 16, size_n * (num_bits // 2)), + device=b_q_weight.device, + dtype=b_q_weight.dtype, + ) + for e in range(num_experts): + output[e] = ops.awq_marlin_repack(b_q_weight[e], size_k, size_n, num_bits) + return output + + +def moe_sum(input: torch.Tensor, output: torch.Tensor): + ops.moe_sum(input, output) + + +def topk_softmax( + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + token_expert_indicies: torch.Tensor, + gating_output: float, +) -> None: + ops.topk_softmax(topk_weights, topk_ids, token_expert_indicies, gating_output) + + +__all__ = [ + "ScalarType", + "awq_marlin_moe_repack", + "fused_experts", + "fused_marlin_moe", + "fused_moe", + "fused_topk", + "gptq_marlin_moe_repack", + "grouped_topk", + "layers", + "moe_align_block_size", + "moe_sum", + "per_token_group_quant_fp8", + "scalar_types", + "topk_softmax", + "w8a8_block_fp8_matmul", +] diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/_moe_69e0359.abi3.so b/build/torch27-cxx11-cu126-aarch64-linux/moe/_moe_69e0359.abi3.so new file mode 100755 index 0000000000000000000000000000000000000000..b0369afa09bcfb531f1e151ddc0e4e643e55b217 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/_moe_69e0359.abi3.so @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f04e6dbd8c1962486dfe12ca4e763470bb072bc2c51d1ca882966af5b7c08b0 +size 67586696 diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/_ops.py b/build/torch27-cxx11-cu126-aarch64-linux/moe/_ops.py new file mode 100644 index 0000000000000000000000000000000000000000..b33dfcc6afacf69f6989c184c5c3efabc03f18da --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/_ops.py @@ -0,0 +1,9 @@ +import torch +from . import _moe_69e0359 +ops = torch.ops._moe_69e0359 + +def add_op_namespace_prefix(op_name: str): + """ + Prefix op by namespace. + """ + return f"_moe_69e0359::{op_name}" \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..56c1a4e3af0b4a93fff71028d8e04bf73f0abb29 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..d3677bebb82a7f3f19344ef6471626493cf2c5bb --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=1792,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=1792,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..265768fb900ccfe9612b4a0d25973e6618f22a79 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=1792,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "5120": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "9216": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "13312": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "17408": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "25600": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "33792": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "41984": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "50176": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "58368": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..d3be23dfc903ba61d3d4d79c0230952b24d2ead0 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "5120": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "9216": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "13312": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "17408": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "25600": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "33792": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "41984": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "50176": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "58368": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=3072,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=3072,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..589f5d39f31418d5121e7cbb2e6f2894b0a7ed32 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=3072,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "5120": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "9216": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "13312": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "17408": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "25600": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "33792": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "41984": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "50176": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "58368": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=3072,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=3072,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..2c78bfaba7890772bf266721f5577202ea443882 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=3072,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "5120": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "9216": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "13312": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "17408": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "25600": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "33792": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "41984": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "50176": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "58368": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=3072,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=3072,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..4da841e74a79f9589fecac1fa557ea132d34805f --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=3072,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "5120": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "9216": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "13312": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "17408": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "25600": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "33792": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "41984": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "50176": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "58368": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=3584,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=3584,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..200356713c0d0a76e199671c7ec8f10d0e5ee0ac --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=3584,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "5120": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "9216": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "13312": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "17408": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "25600": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "33792": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "41984": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "50176": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "58368": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..e076615ee541a5043556f630ecf0946c4e2c1408 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "5120": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "9216": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "13312": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "17408": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "25600": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "33792": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "41984": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "50176": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "58368": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..ee896554b921040d7810bb6e9368cc200777951d --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "5120": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "9216": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "13312": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "17408": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "25600": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "33792": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "41984": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "50176": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "58368": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..05aed8b1c81492151d128ef251afc510d8cc8ed5 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=1,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "5120": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "9216": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "13312": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "17408": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "25600": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "33792": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "41984": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "50176": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "58368": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=1344,device_name=NVIDIA_A100-SXM4-40GB.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=1344,device_name=NVIDIA_A100-SXM4-40GB.json new file mode 100644 index 0000000000000000000000000000000000000000..9262a74a4a0e1e3789f260a3ef7f6cb9551f3f2b --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=1344,device_name=NVIDIA_A100-SXM4-40GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=1344,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=1344,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..d251f9b5accaec977fc87a0999cd56ee387fc650 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=1344,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=1344,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=1344,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..0ecf814a28a9441e89f892eb3d63dcf8dcb0dd97 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=1344,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=14336,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=14336,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..51ad5b299eb22465fa80530d12bdd5d7a03ce398 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=14336,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=14336,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=14336,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..ee5119182556cf49434c10e56cf04e3baeb26408 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=14336,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=1792,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=1792,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..68793c77b33c4f4b97d0a4b780fcbe8043c799de --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=1792,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "5120": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "9216": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "13312": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "17408": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "25600": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "33792": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "41984": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "50176": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "58368": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..612910720ed9439e56c4af4c03f30fee224fac80 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "5120": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "9216": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "13312": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "17408": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "25600": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "33792": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "41984": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "50176": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "58368": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=2688,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=2688,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..039a10ed127b77836a7f41c03513292613852b30 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=2688,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=2688,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=2688,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..3793fcafee60bc7e8f5f12d601cb3192abfa9ca8 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=2688,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=3072,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=3072,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..51d03d8607122d7b9bc20ba48d8432d62367fa00 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=3072,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=3072,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=3072,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..26f9abd6b789e9dd0f83ec7721fd1bae8aa76bec --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=3072,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "1536": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=3200,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=3200,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..cd0cdbea0c3372674cb610870dd0b30325864549 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=3200,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,130 @@ +{ + "3328": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "768": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "1792": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2560": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "2816": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "3584": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "3840": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1280": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "2304": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=3584,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=3584,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..64be6e6591422aa0f441c3747b6c49850929652e --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=3584,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..0a6a6a73fa45e270f01ba7ebdc6d9d55bf9daad3 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "5120": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "9216": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "13312": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "17408": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "25600": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "33792": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "41984": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "50176": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "58368": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=6400,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=6400,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..ba9041d008507e31ae4179ef2bc863a49c606582 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=6400,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,130 @@ +{ + "3840": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "1792": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "3584": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "2816": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1280": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "768": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "3328": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "2560": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "2304": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..7a7508aab04599cb06641c835d8b0a14f54d0716 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..dbf9a2dd6f048d8adee290961e2aea72035f7615 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..bbb2386046b1135a2cc7ab7cb26c1d0b039bcf3a --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "1536": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=800,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=800,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..57055453aa24c831dad9ac8e37fdab707c63ef91 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=16,N=800,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,130 @@ +{ + "2048": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "1792": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "3328": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "2560": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "768": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "2816": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "2304": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2 + }, + "1280": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "3840": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "3584": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=160,N=192,device_name=NVIDIA_A800-SXM4-80GB.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=160,N=192,device_name=NVIDIA_A800-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..0611620eb33626ac891454637723eabec9aa1183 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=160,N=192,device_name=NVIDIA_A800-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=1024,device_name=AMD_Instinct_MI325X,block_shape=[128,128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=1024,device_name=AMD_Instinct_MI325X,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..43c249d2530e1fafb46f6594b73e11f33fa8ffe7 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=1024,device_name=AMD_Instinct_MI325X,block_shape=[128,128].json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=1024,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=1024,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..43c249d2530e1fafb46f6594b73e11f33fa8ffe7 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=1024,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..4dd00d110e486c0a7670504ace778fd5995de727 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..48f9697af2639bd4afd45c768c7a8e9442924e2d --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..a8c05712ba5872db5414e571fef659572e59eb80 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..f1244c61efb019183e892f848700072c3be03ef9 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..2e692a1583a4ad95f020634d55f7f4e327be5da0 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128,128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..a2ee05da1d7c632aa6b1c3ccc8af2546f987ee93 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..63e118746fd86305fd03064637ad915b72c8abcd --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..e6769604ee6bcc75c41a5bea296cb55da3c755e8 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..e6769604ee6bcc75c41a5bea296cb55da3c755e8 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..fc573cd6e8561f56830c8f8ca3033f2d2f7bc1b3 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128,128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..c6d7e96c7f0aea61a2bd57723ecb91b1c1230895 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128,128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..6fcf408755f5d3e28ff636472f89e4fa52c8da22 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_L20,dtype=fp8_w8a8,block_shape=[128,128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_L20,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..c6eabea66a39aa86f5becc04d32ee01dbe627dce --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_L20,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=512,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=512,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..e6769604ee6bcc75c41a5bea296cb55da3c755e8 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=512,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=64,device_name=NVIDIA_A800-SXM4-80GB.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=64,device_name=NVIDIA_A800-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..21f60229ff875ca6dc0cd0a3cb4fa51a845bf881 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=256,N=64,device_name=NVIDIA_A800-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=60,N=1408,device_name=AMD_Instinct_MI300X.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=60,N=1408,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..d09508b31729ddbafc8b5e1dfa88343c61e83ca9 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=60,N=1408,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "8": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=60,N=176,device_name=AMD_Instinct_MI300X.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=60,N=176,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..746463af4d56315af3c31348c2253e43c926553e --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=60,N=176,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "3072": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=60,N=352,device_name=AMD_Instinct_MI300X.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=60,N=352,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..bbdb9ad096451d6c5b9d95fae4ed02ce37715002 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=60,N=352,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "2048": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "3072": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=60,N=704,device_name=AMD_Instinct_MI300X.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=60,N=704,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..43584b1eb6b65f02fc919490b0ff4b9569de58c4 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=60,N=704,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..8cc6c643f236d2f7f9ad29354d9e469d00b20d3f --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_A800-SXM4-80GB.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_A800-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..39a9912fa4bddc48de38d5a8b4fadd71326485d2 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_A800-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..05b54639d234e20096744915c932bb2d415641a5 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..d4c9ddd12972ac0b5fd2be11a9cd1075906e3978 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..c17a4ec346915af69f10f65410450cbd66614810 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H200.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H200.json new file mode 100644 index 0000000000000000000000000000000000000000..170ae7f3fff1d09961d368853dcfc2f24c1b878c --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H200.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=2560,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=2560,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..1d9d352edebc3f228353873677d2369e6a721158 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=2560,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=2560,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=2560,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..9ad5b316750051e4a5dc39b195f6a29eeaa882f0 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=2560,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=2560,device_name=NVIDIA_H200.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=2560,device_name=NVIDIA_H200.json new file mode 100644 index 0000000000000000000000000000000000000000..2883dfd11e7f3a7db429ae0aaeb98d554f271cdf --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=2560,device_name=NVIDIA_H200.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..8abfd84a776b7be2c0920ca8d3a69c06735a08ee --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..2fc18a5e43d298ddabc57dc9f357296890fb7e29 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..be8d4a7fd23d994ac838d45bf249e39764552ccf --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H200.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H200.json new file mode 100644 index 0000000000000000000000000000000000000000..71fdd88643c6ff515808fdc2d67e9fda3ac62ae9 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H200.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..b2799ed3a866e25b78d60d92910c000ebb21ff71 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_A800-SXM4-80GB.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_A800-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..c02de2f628b716456373978f600f9fc8c4b98d24 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_A800-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_GeForce_RTX_4090,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_GeForce_RTX_4090,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..3e0bc75ff87c43ad3a0b3f132e5367a8ef39619a --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_GeForce_RTX_4090,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..9f7ed6726f44e60dea5af588e441b269b6c5cc3d --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..b8d3be2313fa14025d8aeb2fd11e0d1ee997ffa6 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..21b72557e365d01011f48a046915765e67dbd13f --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H200.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H200.json new file mode 100644 index 0000000000000000000000000000000000000000..eaf32f6d76c0a6aa50a5fbc2f055a5e7c9e10b66 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H200.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..b6f1d01f886524c58932ac28c31dcf6f10788e14 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..4bf775347eccb3e76e20ea0018d71343f9848313 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..f245285bd821717f2280fa89b5f0aa6bbce2262d --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI325X.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI325X.json new file mode 100644 index 0000000000000000000000000000000000000000..3918c93b160a39e55153cbf6ff634d59e724d021 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI325X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..3f3ccdafa88f3452a695efad4cb9622d6ae79e6a --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,138 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..841044a4fc6e2da81f18b2765fa30b9de7092b3d --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=NVIDIA_H200.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=NVIDIA_H200.json new file mode 100644 index 0000000000000000000000000000000000000000..59be497fc428755f010e490277f4b945d6ed0195 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=14336,device_name=NVIDIA_H200.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..0e5fd1eec77d7e0aca0f1e3376b2c45217ce234a --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI300X.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..d6ad63509f1573d25fbdc370b93bc39a0d00b51a --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..16e0a91baf31d9a819f27eeea719e564f399379c --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI325X.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI325X.json new file mode 100644 index 0000000000000000000000000000000000000000..d766fc062ddcce10d38bdc68ab1f700d193b8ba7 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI325X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..8323f512db0158570427a7eb0a69bd889bc2d9ed --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI300X.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..1b46cb5716514f290c6b4bc3a18dba490c799272 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..6d5b1ae5b15fc333d3cdaa5dbcacab33267339cc --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI325X.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI325X.json new file mode 100644 index 0000000000000000000000000000000000000000..ffc1b23ea90ddefd24b4b98a222e67e84442cf9c --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI325X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_A100-SXM4-40GB.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_A100-SXM4-40GB.json new file mode 100644 index 0000000000000000000000000000000000000000..f4c0f8417b384870050a95e0cf57edbdf6352b23 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_A100-SXM4-40GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..5c8185cfdeec167ec4b88de51b4b395e28769cc5 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..97c9f4445b166657ad29f1db9fc8281f9c463ec4 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..e4110a5d2e70f4d2698d4c28b1668f8e4a48688a --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_H200.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_H200.json new file mode 100644 index 0000000000000000000000000000000000000000..0883ef40582ea4f1b08c629edcd6dd9beaaf2dad --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_H200.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..81bb765d30031ca3031636d7a442dc2c2292a6a9 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI300X.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..811c77ab4109319c9f079e414bd4f7d2cf4ac9f1 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..2758e48fc4065cd05ceb6fe22574acf2b3aae633 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI325X.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI325X.json new file mode 100644 index 0000000000000000000000000000000000000000..fc31215cbae8afeee27d923e176a25daf1f2b43d --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI325X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..0bb423b28f5ab3825929a4870b96393262a9dd9f --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..55571873395464a3b58f549523905f439a8f1716 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..26bcbf26970c7a77c99e2c8eacd83eefa86967bf --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..1a0aa331933291f0c32b2cf26672771b5c7b4196 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H200.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H200.json new file mode 100644 index 0000000000000000000000000000000000000000..9952be6ba4abe434e2924e9505e834de4ce0528e --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H200.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..379ca107a9469a5e5777e4e648f3f25ffc382a0a --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..5a3f415d54142383fa7677d8192f7a4c4df66b2e --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..6cb80f48329f5718e892fd6f78000df4003832f2 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI325X.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI325X.json new file mode 100644 index 0000000000000000000000000000000000000000..de9d0aba75a7ed45fdfd92c123736f1e981a06eb --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI325X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_A100-SXM4-40GB.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_A100-SXM4-40GB.json new file mode 100644 index 0000000000000000000000000000000000000000..b41f9d443e50678334f906b44fce6d018d69500e --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_A100-SXM4-40GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..edf2a38d12ad3f420f232d2cd61ab149ad138725 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_GeForce_RTX_4090,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_GeForce_RTX_4090,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..32bbadbb9eae8d36e30895d863a8d10ff65167b2 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_GeForce_RTX_4090,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..673bae2ba8ef80ed4d4930739ca7daf0e8f28ee1 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..b2100cebb7f589747430be9ca8c8db368c152d78 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..e6f753cdba35b55e8d021597dc4911bab27caa8d --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H200.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H200.json new file mode 100644 index 0000000000000000000000000000000000000000..53f3394693f060daebff9960a95ed3753e8522c9 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H200.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_L40S.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_L40S.json new file mode 100644 index 0000000000000000000000000000000000000000..d720deb4bdd73d194b1023c99e190b8fcfecdaef --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_L40S.json @@ -0,0 +1,173 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 2 + }, + "2": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 2, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 7 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 128, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 1 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 1 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 2, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 2, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 2, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 2 + }, + "192": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_ctas": 1, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 128, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 8 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_ctas": 1, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_ctas": 1, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 2 + }, + "6144": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 2 + }, + "8192": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_ctas": 1, + "num_stages": 2 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..48bb5f2ccb8e3c623593538461d8d8ba71f93a69 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI300X.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..a64d06c6d1724f736c5081a8b947fa2e6127701d --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..2c49f359c22ac8d3b8fd3b00035b70472c34ed51 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI325X.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI325X.json new file mode 100644 index 0000000000000000000000000000000000000000..c7db6c0cbd3faf615074fc8055ff744e528dd204 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI325X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..dbc624731f5cb9afcdc9213183d00d1e5edd4a00 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..cc614e635ea57327c610ce79e99ae5339614f22e --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..32c0c9da471cbe479044095e0ed14a0f54b73620 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..4dd475c02a19b380efbe8f59945a0f2a41acda11 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H200.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H200.json new file mode 100644 index 0000000000000000000000000000000000000000..2ed15f30fe603c080a2e8a32b664400a50d2153e --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H200.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..bd2c6fbc1b941cebfe8e03f62e3044f5878db0ed --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..8d7b780271854c0fbae29e92cfd3cce508ca029e --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..7a07bbf4141952c4dc6199676b1ef97d1cfabece --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI325X.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI325X.json new file mode 100644 index 0000000000000000000000000000000000000000..3a3268cc17a84d88f0a67e20e266491ad45d1777 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI325X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..f578c8d0160ac3ef85b53c8539d3675455a97173 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..918f6839620cbab1f30b0f9383a9129c2cf2cf3d --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..e341a67917d5177bacb3f6767e7b6d92539826ad --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..eb817268d41208a661d2ab9526c5108d36d44085 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H200.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H200.json new file mode 100644 index 0000000000000000000000000000000000000000..0c7062aea6c4e920cdfdc9e13fc2c0fa38aceb0e --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H200.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..cd4fb8f11b935e401103aab5117a1d1be1799c6e --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI300X.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..cf66868e9d57a541fe10fa9fff682f7a7c6dac97 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..c27ca0a36594fd12a9995fffda60a9d5003425c0 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI325X.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI325X.json new file mode 100644 index 0000000000000000000000000000000000000000..da477b1fb15eb2064cbff78b5e47ae9c54dae7a5 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI325X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..34b916e574f88c65db1dac5889d74a990dc25e9b --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..96cbc111c7fffe8b0a67c22711f9a5c8f28d8b61 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/E=8,N=8192,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=1536,K=1536,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=1536,K=1536,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..6496a38fba8ae09b3025a75f357815b9d6a5e3f4 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=1536,K=1536,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=1536,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=1536,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..3618053b65831b95c4bb0f20ef3b9aa816b2d637 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=1536,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=1536,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=1536,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..46a982f5ee9a4bd67ce244b101c576efeeb53b78 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=1536,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=2048,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=2048,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..035ec027fa56622196b24a03a5042ce010deaebf --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=2048,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=2048,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=2048,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..8b49f2781cb54d19a2789767ebb7e8c3fb55b981 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=2048,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=2304,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=2304,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..851bc9f9f0b50b41451b929eaa518869b6a05412 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=2304,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=2304,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=2304,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..d1227c2157990216d2ca51c69ad0944017f53b6a --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=2304,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=24576,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=24576,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..77ba0d7477bdbcb036a43263e7aaa6b6913f8f4e --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=24576,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=24576,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=24576,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..1c61451fb34e52deec827f8f63c80fb15830c202 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=24576,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=24576,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=24576,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..63e661c80de6a7b1422f7a994a2ee7a4b724911c --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=24576,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=256,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=256,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..cf354037903c0d1fcd077c4647aabce026a723fb --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=256,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=3072,K=1536,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=3072,K=1536,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..0a5d7bfdba4852da9ed08d1bc27cd7d521d09965 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=3072,K=1536,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=3072,K=1536,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=3072,K=1536,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..eccb86a76df0d7302b760ab6d83a8ceb9fa9d0d9 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=3072,K=1536,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=3072,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=3072,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..cb91a279d423d0ca25197e0edd5e8c2f4da58720 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=3072,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=3072,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=3072,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..88af48431d8b8791af8df03429704606b670f1f7 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=3072,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=3072,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=3072,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..dd069726d7ed4dcbb449af243f4f4af21815f854 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=3072,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=32768,K=512,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=32768,K=512,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..7febe3d272b4bb76500f7c6b523396129fd53680 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=32768,K=512,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=32768,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=32768,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..56b939e52fac3ed53a4e0ba640c40010cb3af30a --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=32768,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=32768,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=32768,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..63d9a0bf5d79ddaaad547d44338ad4b959ad72b1 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=32768,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=36864,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=36864,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..7fa398c15a2a535401709b0f25e20f6e4b23e58e --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=36864,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=36864,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=36864,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..f15d8f64c7090bd71d0091a524c65d7818fec38e --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=36864,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=4096,K=512,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=4096,K=512,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..9d7658bfc41b2c8fd4daf3fbdf62d15936d3d546 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=4096,K=512,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=4096,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=4096,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..cd3e07804fdec10c2cfb291c1ede3ba67b753f9c --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=4096,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=4096,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=4096,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..9d5a329d7466a37c0ca68a65a089fbb99f9327a9 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=4096,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=4608,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=4608,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..03dba5ad15ba5f7f49100a5c78e8685e64334b2a --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=4608,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=4608,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=4608,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..96e1594a3eabbaedc792b84b07f05ae8752b7251 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=4608,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=4608,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=4608,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..5ffd367df833d773355590220598a3c7eceba4e0 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=4608,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=512,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=512,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..9a5ff48b8942957dde9b862aed848390dd267948 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=512,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=512,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=512,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..eabc423949a24c2a1fb2368a73e5249caf8d07df --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=512,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=576,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=576,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..386928de139ce718f28222b9c1a6555df3958491 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=576,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 5 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=576,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=576,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..51e237b91b8e775a36bcf783c078c2c1cecbcbd2 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=576,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=576,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=576,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..6280219c9ee7d26f7e2fd3625dc92d847ddc7982 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=576,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=1024,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=1024,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..40c01c0b92b4b26fe480879dda33f18c5eb59a6d --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=1024,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=1024,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=1024,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..c6fd3659799bc31e17f3577e7f0e8d7268faf1fb --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=1024,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=1152,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=1152,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..160f12ed3f95a6967439ff53bc3e3a2cdc97c700 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=1152,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=1152,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=1152,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..e5c4a1d2c94e5c7864f462e083ea5f530b8efe3f --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=1152,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=128,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=128,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..2bf5eb27e38208871d50348b170c8c74b80fc519 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=128,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "96": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=16384,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=16384,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..9c908e80406587da4d246ce4e3a8a98a14c875b1 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=16384,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=16384,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=16384,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..0a1e14cffbb2a894a701352193947d272427db0d --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=16384,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=16384,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=16384,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..15b1c93f60fc5068ba11b82b6d5924dd2024a824 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=16384,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=18432,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=18432,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..f78e7060e6840ff721d306db556636b0bbc8d9b3 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=18432,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=18432,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=18432,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..8ff12e64c172f5a5d0fbdf900728fe60b33877e2 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=18432,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=18432,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=18432,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..4532f93681e2be175b1bf94f81bfde711821cd60 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=18432,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=2048,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=2048,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..1d3ce5c94c2d9a4a1637204efb3b14f7a5579bdb --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=2048,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=2048,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=2048,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..ca7f32b9552b479dc05495792b7e426db5eb1b56 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=2048,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=2048,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=2048,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..5acea242cc0ad094cba8ee5f568ff88afb1b41ae --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=2048,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=2304,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=2304,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..3ab5796ee15b6ec8d4ab1f4ab5a594fecb30e4b4 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=2304,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=2304,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=2304,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..58cdd93e90b8c29bc7a211861711565dbeeb529a --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=2304,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=2304,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=2304,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..b72e0371d1421a1decc9d57860f83eea8f790942 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=2304,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=256,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=256,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..3cb7eaa07c745fd3aa2b3242780a7061bedac1de --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=256,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..293adce387e066fce75b6e606d4b8b6a5aa10bdb --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/configs/N=7168,K=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + } +} diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/fp8.py b/build/torch27-cxx11-cu126-aarch64-linux/moe/fp8.py new file mode 100644 index 0000000000000000000000000000000000000000..23bd7d6703104b0020671cc6ba6f78a6df37e4bf --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/fp8.py @@ -0,0 +1,63 @@ +from typing import Optional, Tuple, Union + +import torch + +from ._ops import ops + + +def is_hip() -> bool: + return torch.version.hip is not None + + +def scaled_fp8_quant( + input: torch.Tensor, + scale: Optional[torch.Tensor] = None, + num_token_padding: Optional[int] = None, + scale_ub: Optional[torch.Tensor] = None, + use_per_token_if_dynamic: bool = False, +) -> Tuple[torch.Tensor, torch.Tensor]: + """ + Quantize input tensor to FP8 and return quantized tensor and scale. + + This function supports both static and dynamic quantization: If you + provide the scale, it will use static scaling and if you omit it, + the scale will be determined dynamically. The function also allows + optional padding of the output tensors for downstream kernels that + will benefit from padding. + + Args: + input: The input tensor to be quantized to FP8 + scale: Optional scaling factor for the FP8 quantization + scale_ub: Optional upper bound for scaling factor in dynamic + per token case + num_token_padding: If specified, pad the first dimension + of the output to at least this value. + use_per_token_if_dynamic: Whether to do per_tensor or per_token + in the dynamic quantization case. + + Returns: + Tuple[torch.Tensor, torch.Tensor]: The output tensor in FP8 and + scaling factor. + """ + # This code assumes batch_dim and num_tokens are flattened + assert input.ndim == 2 + shape: Union[Tuple[int, int], torch.Size] = input.shape + # For rocm, the output fp8 dtype is torch.float_e3m3fnuz + out_dtype: torch.dtype = torch.float8_e4m3fnuz if is_hip() else torch.float8_e4m3fn + if num_token_padding: + shape = (max(num_token_padding, input.shape[0]), shape[1]) + output = torch.empty(shape, device=input.device, dtype=out_dtype) + + if scale is None: + if use_per_token_if_dynamic: + scale = torch.empty((shape[0], 1), device=input.device, dtype=torch.float32) + ops.dynamic_per_token_scaled_fp8_quant(output, input, scale, scale_ub) + else: + scale = torch.zeros(1, device=input.device, dtype=torch.float32) + ops.dynamic_scaled_fp8_quant(output, input, scale) + else: + # num_token_padding not implemented for this case + assert scale.numel() == 1 or num_token_padding is None + ops.static_scaled_fp8_quant(output, input, scale) + + return output, scale diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/fp8_utils.py b/build/torch27-cxx11-cu126-aarch64-linux/moe/fp8_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..acb4f3e3bb1a34f209fdac9ecca8c123aaf67f12 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/fp8_utils.py @@ -0,0 +1,391 @@ +import functools +import json +import logging +import os +from typing import Any, Dict, List, Optional, Tuple + +import torch +import triton +import triton.language as tl + +from .platforms import current_platform + +logger = logging.getLogger(__name__) + + +@triton.jit +def _per_token_group_quant_fp8( + # Pointers to inputs and output + y_ptr, + y_q_ptr, + y_s_ptr, + group_size, + # Avoid to divide zero + eps, + # Information for float8 + fp8_min, + fp8_max, + # Meta-parameters + BLOCK: tl.constexpr, +): + """A Triton-accelerated function to perform per-token-group + quantization on a tensor. + This function converts the tensor values into float8 values. + """ + # Map the program id to the row of X and Y it should compute. + g_id = tl.program_id(0) + y_ptr += g_id * group_size + y_q_ptr += g_id * group_size + y_s_ptr += g_id + + cols = tl.arange(0, BLOCK) # N <= BLOCK + mask = cols < group_size + + y = tl.load(y_ptr + cols, mask=mask, other=0.0).to(tl.float32) + # Quant + _absmax = tl.maximum(tl.max(tl.abs(y)), eps) + y_s = _absmax / fp8_max + y_q = tl.clamp(y / y_s, fp8_min, fp8_max).to(y_q_ptr.dtype.element_ty) + + tl.store(y_q_ptr + cols, y_q, mask=mask) + tl.store(y_s_ptr, y_s) + + +@triton.jit +def _per_token_group_quant_fp8_colmajor( + # Pointers to inputs and output + y_ptr, + y_q_ptr, + y_s_ptr, + group_size, + # Num columns of y + y_num_columns, + # Stride from one column to the next of y_s + y_s_col_stride, + # Avoid to divide zero + eps, + # Information for float8 + fp8_min, + fp8_max, + # Meta-parameters + BLOCK: tl.constexpr, +): + """A Triton-accelerated function to perform per-token-group + quantization on a tensor. + This function converts the tensor values into float8 values. + """ + # Map the program id to the row of X and Y it should compute. + g_id = tl.program_id(0) + y_ptr += g_id * group_size + y_q_ptr += g_id * group_size + + # Convert g_id the flattened block coordinate to 2D so we can index + # into the output y_scales matrix + blocks_per_row = y_num_columns // group_size + scale_col = g_id % blocks_per_row + scale_row = g_id // blocks_per_row + y_s_ptr += scale_col * y_s_col_stride + scale_row + + cols = tl.arange(0, BLOCK) # group_size <= BLOCK + mask = cols < group_size + + y = tl.load(y_ptr + cols, mask=mask, other=0.0).to(tl.float32) + # Quant + _absmax = tl.maximum(tl.max(tl.abs(y)), eps) + y_s = _absmax / fp8_max + y_q = tl.clamp(y / y_s, fp8_min, fp8_max).to(y_q_ptr.dtype.element_ty) + + tl.store(y_q_ptr + cols, y_q, mask=mask) + tl.store(y_s_ptr, y_s) + + +def per_token_group_quant_fp8( + x: torch.Tensor, + group_size: int, + eps: float = 1e-10, + dtype: Optional[torch.dtype] = None, + column_major_scales: bool = False, +) -> Tuple[torch.Tensor, torch.Tensor]: + """Function to perform per-token-group quantization on an input tensor `x`. + It converts the tensor values into signed float8 values and returns the + quantized tensor along with the scaling factor used for quantization. + Args: + x: The input tensor with ndim >= 2. + group_size: The group size used for quantization. + eps: The minimum to avoid dividing zero. + dtype: The dype of output tensor. Note that only `torch.float8_e4m3fn` + is supported for now. + Returns: + Tuple[torch.Tensor, torch.Tensor]: The quantized tensor and the + scaling factor for quantization. + """ + if dtype is None: + dtype = ( + torch.float8_e4m3fnuz if current_platform.is_rocm() else torch.float8_e4m3fn + ) + assert x.shape[-1] % group_size == 0, ( + f"the last dimension of `x` {x.shape[-1]} must be divisible " + f"by `group_size` {group_size}" + ) + assert x.is_contiguous(), "`x` must be contiguous" + + finfo = torch.finfo(dtype) + fp8_min = finfo.min + fp8_max = finfo.max + + x_q = torch.empty_like(x, device=x.device, dtype=dtype) + M = x.numel() // group_size + N = group_size + if column_major_scales: + shape = (x.shape[-1] // group_size,) + x.shape[:-1] + x_s = torch.empty(shape, device=x.device, dtype=torch.float32).permute(-1, -2) + else: + shape = x.shape[:-1] + (x.shape[-1] // group_size,) + x_s = torch.empty(shape, device=x.device, dtype=torch.float32) + + BLOCK = triton.next_power_of_2(N) + # heuristics for number of warps + num_warps = min(max(BLOCK // 256, 1), 8) + num_stages = 1 + if column_major_scales: + _per_token_group_quant_fp8_colmajor[(M,)]( + x, + x_q, + x_s, + group_size, + x.shape[1], + x_s.stride(1), + eps, + fp8_min=fp8_min, + fp8_max=fp8_max, + BLOCK=BLOCK, + num_warps=num_warps, + num_stages=num_stages, + ) + else: + _per_token_group_quant_fp8[(M,)]( + x, + x_q, + x_s, + group_size, + eps, + fp8_min=fp8_min, + fp8_max=fp8_max, + BLOCK=BLOCK, + num_warps=num_warps, + num_stages=num_stages, + ) + + return x_q, x_s + + +@triton.jit +def _w8a8_block_fp8_matmul( + # Pointers to inputs and output + A, + B, + C, + As, + Bs, + # Shape for matmul + M, + N, + K, + # Block size for block-wise quantization + group_n, + group_k, + # Stride for inputs and output + stride_am, + stride_ak, + stride_bk, + stride_bn, + stride_cm, + stride_cn, + stride_As_m, + stride_As_k, + stride_Bs_k, + stride_Bs_n, + # Meta-parameters + BLOCK_SIZE_M: tl.constexpr, + BLOCK_SIZE_N: tl.constexpr, + BLOCK_SIZE_K: tl.constexpr, + GROUP_SIZE_M: tl.constexpr, +): + """Triton-accelerated function used to perform linear operations (dot + product) on input tensors `A` and `B` with block-wise quantization, and + store the result in output tensor `C`. + """ + + pid = tl.program_id(axis=0) + num_pid_m = tl.cdiv(M, BLOCK_SIZE_M) + num_pid_n = tl.cdiv(N, BLOCK_SIZE_N) + num_pid_in_group = GROUP_SIZE_M * num_pid_n + group_id = pid // num_pid_in_group + first_pid_m = group_id * GROUP_SIZE_M + group_size_m = min(num_pid_m - first_pid_m, GROUP_SIZE_M) + pid_m = first_pid_m + (pid % group_size_m) + pid_n = (pid % num_pid_in_group) // group_size_m + + offs_am = (pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)) % M + offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)) % N + offs_k = tl.arange(0, BLOCK_SIZE_K) + a_ptrs = A + (offs_am[:, None] * stride_am + offs_k[None, :] * stride_ak) + b_ptrs = B + (offs_k[:, None] * stride_bk + offs_bn[None, :] * stride_bn) + + As_ptrs = As + offs_am * stride_As_m + offs_bsn = offs_bn // group_n + Bs_ptrs = Bs + offs_bsn * stride_Bs_n + + accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32) + for k in range(0, tl.cdiv(K, BLOCK_SIZE_K)): + a = tl.load(a_ptrs, mask=offs_k[None, :] < K - k * BLOCK_SIZE_K, other=0.0) + b = tl.load(b_ptrs, mask=offs_k[:, None] < K - k * BLOCK_SIZE_K, other=0.0) + + k_start = k * BLOCK_SIZE_K + offs_ks = k_start // group_k + a_s = tl.load(As_ptrs + offs_ks * stride_As_k) + b_s = tl.load(Bs_ptrs + offs_ks * stride_Bs_k) + + accumulator += tl.dot(a, b) * a_s[:, None] * b_s[None, :] + a_ptrs += BLOCK_SIZE_K * stride_ak + b_ptrs += BLOCK_SIZE_K * stride_bk + + if C.dtype.element_ty == tl.bfloat16: + c = accumulator.to(tl.bfloat16) + elif C.dtype.element_ty == tl.float16: + c = accumulator.to(tl.float16) + else: + c = accumulator.to(tl.float32) + + offs_cm = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M) + offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N) + c_ptrs = C + stride_cm * offs_cm[:, None] + stride_cn * offs_cn[None, :] + c_mask = (offs_cm[:, None] < M) & (offs_cn[None, :] < N) + tl.store(c_ptrs, c, mask=c_mask) + + +@functools.lru_cache +def get_w8a8_block_fp8_configs( + N: int, K: int, block_n: int, block_k: int +) -> Optional[Dict[int, Any]]: + """ + Return optimized configurations for the w8a8 block fp8 kernel. + The return value will be a dictionary that maps an irregular grid of + batch sizes to configurations of the w8a8 block fp8 kernel. To evaluate the + kernel on a given batch size bs, the closest batch size in the grid should + be picked and the associated configuration chosen to invoke the kernel. + """ + + # First look up if an optimized configuration is available in the configs + # directory + device_name = current_platform.get_device_name().replace(" ", "_") + json_file_name = f"N={N},K={K},device_name={device_name},dtype=fp8_w8a8,block_shape=[{block_n}, {block_k}].json" # noqa: E501 + + config_file_path = os.path.join( + os.path.dirname(os.path.realpath(__file__)), "configs", json_file_name + ) + if os.path.exists(config_file_path): + with open(config_file_path) as f: + logger.info( + "Using configuration from %s for W8A8 Block FP8 kernel.", + config_file_path, + ) + # If a configuration has been found, return it + return {int(key): val for key, val in json.load(f).items()} + + # If no optimized configuration is available, we will use the default + # configuration + logger.warning( + "Using default W8A8 Block FP8 kernel config. Performance might " + "be sub-optimal! Config file not found at %s", + config_file_path, + ) + return None + + +def w8a8_block_fp8_matmul( + A: torch.Tensor, + B: torch.Tensor, + As: torch.Tensor, + Bs: torch.Tensor, + block_size: List[int], + output_dtype: torch.dtype = torch.float16, +) -> torch.Tensor: + """This function performs matrix multiplication with block-wise + quantization. + It takes two input tensors `A` and `B` with scales `As` and `Bs`. + The output is returned in the specified `output_dtype`. + Args: + A: The input tensor, e.g., activation. + B: The input tensor, e.g., weight. + As: The per-token-group quantization scale for `A`. + Bs: The per-block quantization scale for `B`. + block_size: The block size for per-block quantization. It should + be 2-dim, e.g., [128, 128]. + output_dytpe: The dtype of the returned tensor. + Returns: + torch.Tensor: The result of matmul. + """ + assert len(block_size) == 2 + block_n, block_k = block_size[0], block_size[1] + + assert A.shape[-1] == B.shape[-1] + assert A.shape[:-1] == As.shape[:-1] and A.is_contiguous() + assert triton.cdiv(A.shape[-1], block_k) == As.shape[-1] + M = A.numel() // A.shape[-1] + + assert B.ndim == 2 and B.is_contiguous() and Bs.ndim == 2 + N, K = B.shape + assert triton.cdiv(N, block_n) == Bs.shape[0] + assert triton.cdiv(K, block_k) == Bs.shape[1] + + C_shape = A.shape[:-1] + (N,) + C = A.new_empty(C_shape, dtype=output_dtype) + + configs = get_w8a8_block_fp8_configs(N, K, block_size[0], block_size[1]) + if configs: + # Get the optimal config if there is one + config = configs[min(configs.keys(), key=lambda x: abs(x - M))] + else: + # Default config + # Block-wise quant: BLOCK_SIZE_N must be divisible by block_size[0] + # BLOCK_SIZE_K must be divisible by block_size[1] + config = { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": block_size[0], + "BLOCK_SIZE_K": block_size[1], + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2, + } + + def grid(META): + return ( + triton.cdiv(M, META["BLOCK_SIZE_M"]) * triton.cdiv(N, META["BLOCK_SIZE_N"]), + ) + + _w8a8_block_fp8_matmul[grid]( + A, + B, + C, + As, + Bs, + M, + N, + K, + block_n, + block_k, + A.stride(-2), + A.stride(-1), + B.stride(1), + B.stride(0), + C.stride(-2), + C.stride(-1), + As.stride(-2), + As.stride(-1), + Bs.stride(1), + Bs.stride(0), + **config, + ) + + return C diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/fused_marlin_moe.py b/build/torch27-cxx11-cu126-aarch64-linux/moe/fused_marlin_moe.py new file mode 100644 index 0000000000000000000000000000000000000000..b3e0a5c24599730faf973fad3cf3fb6031a30522 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/fused_marlin_moe.py @@ -0,0 +1,413 @@ +"""Fused MoE utilities for GPTQ.""" + +import functools +from typing import TYPE_CHECKING, Optional + +import torch + +from ._ops import add_op_namespace_prefix, ops +from .fused_moe import fused_topk, moe_align_block_size, try_get_optimal_moe_config +from .scalar_type import ScalarType, scalar_types + +# neuron has torch version that doesn't even have impl_abstract +if TYPE_CHECKING: + + def register_fake(fn): + return lambda name: fn + +else: + try: + from torch.library import register_fake + except ImportError: + from torch.library import impl_abstract as register_fake + + +def get_scalar_type(num_bits: int, has_zp: bool): + if has_zp: + assert num_bits == 4 + return scalar_types.uint4 + else: + return scalar_types.uint4b8 if num_bits == 4 else scalar_types.uint8b128 + + +def single_marlin_moe( + hidden_states: torch.Tensor, + w: torch.Tensor, + scales: torch.Tensor, + gating_output: torch.Tensor, + topk: int, + renormalize: bool, + g_idx: Optional[torch.Tensor] = None, + sort_indices: Optional[torch.Tensor] = None, + w_zeros: Optional[torch.Tensor] = None, + num_bits: int = 8, + is_k_full: bool = True, +) -> torch.Tensor: + """ + This function computes the multiplication of hidden_states with expert + weights used in Marlin MoE, using weights w and top-k gating mechanism. + Its purpose is testing and debugging the fused MoE kernel. + + Parameters: + - hidden_states (torch.Tensor): The input tensor to the Marlin Mul. + - w (torch.Tensor): The set of expert weights. + - scales (torch.Tensor): The quantization scales. + - gating_output (torch.Tensor): The output of the gating operation + (before softmax). + - g_idx (Optional[torch.Tensor]): Optional act_order indices. + - sort_indices (Optional[torch.Tensor]): Optional act_order input + permutation. + - topk (int): The number of top-k experts to select. + - renormalize (bool): If True, renormalize the top-k weights to sum to 1. + - w_zeros (Optional[torch.Tensor]): Optional zero points to be used for w. + - num_bits (bool): The number of bits in expert weights quantization. + + Returns: + - torch.Tensor: The output tensor after applying the MoE layer. + """ + # Check constraints. + assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch" + assert hidden_states.shape[1] == w.shape[1] * 16, "Hidden size mismatch" + assert gating_output.shape[1] == w.shape[0], "Number of experts mismatch" + assert hidden_states.is_contiguous(), "Hidden_states must be contiguous" + assert w.is_contiguous(), "Expert weights must be contiguous" + assert hidden_states.dtype == torch.float16 + assert num_bits in [4, 8] + + M, K = hidden_states.shape + E = w.shape[0] + N = w.shape[2] // (num_bits // 2) + + topk_weights, topk_ids = fused_topk(hidden_states, gating_output, topk, renormalize) + + # This might not be an optimal config for a single MMM + get_config_func = functools.partial( + try_get_optimal_moe_config, + w.shape, + w.shape, + topk_ids.shape[1], + None, + is_marlin=True, + ) + config = get_config_func(M) + + block_size_m = config["BLOCK_SIZE_M"] + + sorted_token_ids, _, _ = moe_align_block_size(topk_ids, block_size_m, E) + + max_workspace_size = (N // 64) * 16 + workspace = torch.zeros( + max_workspace_size, + dtype=torch.int, + device=hidden_states.device, + requires_grad=False, + ) + + has_zero_point = w_zeros is not None + if w_zeros is None: + w_zeros = torch.empty( + (0, 0), + dtype=hidden_states.dtype, + device=hidden_states.device, + requires_grad=False, + ) + + if g_idx is None: + g_idx = torch.empty( + (0, 0), dtype=torch.int32, device=hidden_states.device, requires_grad=False + ) + + if sort_indices is None: + sort_indices = torch.empty( + (0), dtype=torch.int32, device=hidden_states.device, requires_grad=False + ) + + scalar_type = get_scalar_type(num_bits, has_zero_point) + + intermediate_cache = ops.marlin_gemm_moe( + hidden_states, + w, + sorted_token_ids, + topk_weights, + topk_ids, + scales, + w_zeros, + g_idx, + sort_indices, + workspace, + scalar_type.id, + M, + N, + K, + is_k_full, + E, + topk, + block_size_m, + True, + False, + ) + + return torch.sum(intermediate_cache.view(*intermediate_cache.shape), dim=1) + + +if hasattr(ops, "single_marlin_gemm_moe"): + + @register_fake(add_op_namespace_prefix("single_marlin_gemm_moe")) + def single_marlin_moe_fake( + hidden_states: torch.Tensor, + w: torch.Tensor, + scales: torch.Tensor, + gating_output: torch.Tensor, + topk: int, + renormalize: bool, + g_idx: Optional[torch.Tensor] = None, + sort_indices: Optional[torch.Tensor] = None, + w_zeros: Optional[torch.Tensor] = None, + num_bits: int = 8, + is_k_full: bool = True, + ) -> torch.Tensor: + return torch.empty_like(hidden_states) + + +def fused_marlin_moe( + hidden_states: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + w1_scale: torch.Tensor, + w2_scale: torch.Tensor, + gating_output: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + g_idx1: Optional[torch.Tensor] = None, + g_idx2: Optional[torch.Tensor] = None, + sort_indices1: Optional[torch.Tensor] = None, + sort_indices2: Optional[torch.Tensor] = None, + w1_zeros: Optional[torch.Tensor] = None, + w2_zeros: Optional[torch.Tensor] = None, + num_bits: int = 8, + is_k_full: bool = True, +) -> torch.Tensor: + """ + This function computes a Mixture of Experts (MoE) layer using two sets of + weights, w1 and w2, and top-k gating mechanism. + + Parameters: + - hidden_states (torch.Tensor): The input tensor to the MoE layer. + - w1 (torch.Tensor): The first set of expert weights. + - w2 (torch.Tensor): The second set of expert weights. + - w1_scale (torch.Tensor): Scale to be used for w1. + - w2_scale (torch.Tensor): Scale to be used for w2. + - gating_output (torch.Tensor): The output of the gating operation + (before softmax). + - g_idx1 (Optional[torch.Tensor]): The first set of act_order indices. + - g_idx2 (Optional[torch.Tensor]): The second set of act_order indices. + - sort_indices1 (Optional[torch.Tensor]): The first act_order input + permutation. + - sort_indices2 (Optional[torch.Tensor]): The second act_order input + permutation. + - topk_weights (torch.Tensor): Top-k weights. + - topk_ids (torch.Tensor): Indices of topk-k elements. + - w1_zeros (Optional[torch.Tensor]): Optional zero points to be used for w1. + - w2_zeros (Optional[torch.Tensor]): Optional zero points to be used for w2. + - num_bits (bool): The number of bits in expert weights quantization. + + Returns: + - torch.Tensor: The output tensor after applying the MoE layer. + """ + # Check constraints. + assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch" + assert hidden_states.shape[1] == w1.shape[1] * 16, "Hidden size mismatch w1" + assert hidden_states.shape[1] == w2.shape[2] // ( + num_bits // 2 + ), "Hidden size mismatch w2" + assert gating_output.shape[1] == w1.shape[0], "Number of experts mismatch" + assert hidden_states.is_contiguous(), "Hidden_states must be contiguous" + assert w1.is_contiguous(), "Expert weights1 must be contiguous" + assert w2.is_contiguous(), "Expert weights2 must be contiguous" + assert hidden_states.dtype == torch.float16 + assert num_bits in [4, 8] + + has_no_act_order = ( + g_idx1 is None + and g_idx2 is None + and sort_indices1 is None + and sort_indices2 is None + ) + has_all_act_order = ( + g_idx1 is not None + and g_idx2 is not None + and sort_indices1 is not None + and sort_indices2 is not None + ) + assert has_no_act_order or has_all_act_order, ( + "g_idx and sorted_indices " "must be all not None or must be all None" + ) + + has_no_zp = w1_zeros is None and w2_zeros is None + has_all_zp = w1_zeros is not None and w2_zeros is not None + assert has_no_zp or has_all_zp, ( + "zero points must be both not None or " "must be both None" + ) + + M, K = hidden_states.shape + E = w1.shape[0] + N = w2.shape[1] * 16 + topk = topk_ids.shape[1] + + get_config_func = functools.partial( + try_get_optimal_moe_config, + w1.shape, + w2.shape, + topk_ids.shape[1], + None, + is_marlin=True, + ) + config = get_config_func(M) + + block_size_m = config["BLOCK_SIZE_M"] + + sorted_token_ids, _, _ = moe_align_block_size(topk_ids, block_size_m, E) + + max_workspace_size = (max(2 * N, K) // 64) * 16 + workspace = torch.zeros( + max_workspace_size, dtype=torch.int, device="cuda", requires_grad=False + ) + + if has_no_zp: + w1_zeros = torch.empty( + (0, 0), + dtype=hidden_states.dtype, + device=hidden_states.device, + requires_grad=False, + ) + w2_zeros = torch.empty( + (0, 0), + dtype=hidden_states.dtype, + device=hidden_states.device, + requires_grad=False, + ) + + if has_no_act_order: + g_idx1 = torch.empty( + (0, 0), dtype=torch.int32, device=hidden_states.device, requires_grad=False + ) + g_idx2 = torch.empty( + (0, 0), dtype=torch.int32, device=hidden_states.device, requires_grad=False + ) + sort_indices1 = torch.empty( + (0), dtype=torch.int32, device=hidden_states.device, requires_grad=False + ) + sort_indices2 = torch.empty( + (0, 0), dtype=torch.int32, device=hidden_states.device, requires_grad=False + ) + + scalar_type1 = get_scalar_type(num_bits, has_all_zp) + scalar_type2 = get_scalar_type(num_bits, has_all_zp) + + intermediate_cache2 = torch.empty( + (M * topk_ids.shape[1], N), + device=hidden_states.device, + dtype=hidden_states.dtype, + ) + + intermediate_cache1 = ops.marlin_gemm_moe( + hidden_states, + w1, + sorted_token_ids, + topk_weights, + topk_ids, + w1_scale, + w1_zeros, + g_idx1, + sort_indices1, + workspace, + scalar_type1.id, + M, + 2 * N, + K, + is_k_full, + E, + topk, + block_size_m, + True, + False, + ) + + ops.silu_and_mul(intermediate_cache2, intermediate_cache1.view(-1, 2 * N)) + + intermediate_cache3 = ops.marlin_gemm_moe( + intermediate_cache2, + w2, + sorted_token_ids, + topk_weights, + topk_ids, + w2_scale, + w2_zeros, + g_idx2, + sort_indices2, + workspace, + scalar_type2.id, + M, + K, + N, + is_k_full, + E, + topk, + block_size_m, + False, + True, + ) + + return torch.sum(intermediate_cache3.view(*intermediate_cache3.shape), dim=1) + + +if hasattr(ops, "fused_marlin_moe"): + + @register_fake(add_op_namespace_prefix("fused_marlin_moe")) + def fused_marlin_moe_fake( + hidden_states: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + w1_scale: torch.Tensor, + w2_scale: torch.Tensor, + gating_output: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + g_idx1: Optional[torch.Tensor] = None, + g_idx2: Optional[torch.Tensor] = None, + sort_indices1: Optional[torch.Tensor] = None, + sort_indices2: Optional[torch.Tensor] = None, + w1_zeros: Optional[torch.Tensor] = None, + w2_zeros: Optional[torch.Tensor] = None, + num_bits: int = 8, + is_k_full: bool = True, + ) -> torch.Tensor: + return torch.empty_like(hidden_states) + + +if hasattr(ops, "marlin_gemm_moe"): + + @register_fake(add_op_namespace_prefix("marlin_gemm_moe")) + def marlin_gemm_moe_fake( + a: torch.Tensor, + b_q_weights: torch.Tensor, + sorted_ids: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + b_scales: torch.Tensor, + b_zero_points: torch.Tensor, + g_idx: torch.Tensor, + perm: torch.Tensor, + workspace: torch.Tensor, + b_q_type: ScalarType, + size_m: torch.SymInt, + size_n: torch.SymInt, + size_k: torch.SymInt, + is_k_full: bool, + num_experts: int, + topk: int, + moe_block_size: int, + replicate_input: bool, + apply_weights: bool, + ) -> torch.Tensor: + return torch.empty((size_m, topk, size_n), dtype=a.dtype, device=a.device) diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/fused_moe.py b/build/torch27-cxx11-cu126-aarch64-linux/moe/fused_moe.py new file mode 100644 index 0000000000000000000000000000000000000000..f8f0586c2a58d1fb68f7d2eaed47a48f165d1e8d --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/fused_moe.py @@ -0,0 +1,1574 @@ +# SPDX-License-Identifier: Apache-2.0 +"""Fused MoE kernel.""" + +import functools +import json +import logging +import os +from typing import Any, Callable, Dict, List, Optional, Tuple + +import torch +import triton +import triton.language as tl + +from ._ops import ops +from .fp8 import scaled_fp8_quant +from .fp8_utils import per_token_group_quant_fp8 +from .moe_align_block_size import moe_align_block_size +from .platforms import current_platform +from .utils.custom_ops import direct_register_custom_op + +logger = logging.getLogger(__name__) + + +VLLM_FUSED_MOE_CHUNK_SIZE = int(os.getenv("VLLM_FUSED_MOE_CHUNK_SIZE", "32768")) +VLLM_ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON: bool = bool( + int(os.getenv("VLLM_ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON", "0")) +) + + +def cdiv(a: int, b: int) -> int: + """Ceiling division.""" + return -(a // -b) + + +def _fp8_quantize( + A: torch.Tensor, + A_scale: Optional[torch.Tensor], + block_shape: Optional[List[int]], +) -> Tuple[torch.Tensor, torch.Tensor]: + """ + Perform fp8 quantization on the inputs. If a block_shape + is provided, the output will be blocked. + """ + if block_shape is None: + A, A_scale = scaled_fp8_quant(A, A_scale) + else: + assert len(block_shape) == 2 + _, block_k = block_shape[0], block_shape[1] + A, A_scale = per_token_group_quant_fp8(A, block_k) + assert cdiv(A.shape[-1], block_k) == A_scale.shape[-1] + return A, A_scale + + +@triton.jit +def write_zeros_to_output( + c_ptr, + stride_cm, + stride_cn, + pid_n, + N, + offs_token, + token_mask, + BLOCK_SIZE_M, + BLOCK_SIZE_N, + compute_type, +): + accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=compute_type) + offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N) + c_ptrs = c_ptr + stride_cm * offs_token[:, None] + stride_cn * offs_cn[None, :] + c_mask = token_mask[:, None] & (offs_cn[None, :] < N) + tl.store(c_ptrs, accumulator, mask=c_mask) + + +@triton.jit +def fused_moe_kernel_gptq_awq( + # Pointers to matrices + a_ptr, + b_ptr, + c_ptr, + b_scale_ptr, + b_zp_ptr, + topk_weights_ptr, + sorted_token_ids_ptr, + expert_ids_ptr, + num_tokens_post_padded_ptr, + # Matrix dimensions + N: tl.constexpr, + K: tl.constexpr, + EM, + num_valid_tokens, + # The stride variables represent how much to increase the ptr by when + # moving by 1 element in a particular dimension. E.g. `stride_am` is + # how much to increase `a_ptr` by to get the element one row down + # (A has M rows). + stride_am, + stride_ak, + stride_be, + stride_bk, + stride_bn, + stride_cm, + stride_cn, + stride_bse, + stride_bsk, + stride_bsn, + stride_bze, + stride_bzk, + stride_bzn, + block_k_diviable: tl.constexpr, + group_size: tl.constexpr, + # Meta-parameters + BLOCK_SIZE_M: tl.constexpr, + BLOCK_SIZE_N: tl.constexpr, + BLOCK_SIZE_K: tl.constexpr, + GROUP_SIZE_M: tl.constexpr, + MUL_ROUTED_WEIGHT: tl.constexpr, + top_k: tl.constexpr, + compute_type: tl.constexpr, + has_zp: tl.constexpr, + use_int4_w4a16: tl.constexpr, + use_int8_w8a16: tl.constexpr, +): + """ + Implements the fused computation for a Mixture of Experts (MOE) using + token and expert matrices. + + Key Parameters: + - A: The input tensor representing tokens with shape (*, K), where '*' can + be any shape representing batches and K is the feature dimension of + each token. + - B: The stacked MOE weight tensor with shape (E, N, K), where E is + the number of experts, K is the input feature dimension, and N is + the output feature dimension. + - C: The output cache tensor with shape (M, topk, N), where M is the + total number of tokens post padding, topk is the number of times + each token is repeated, and N is the output feature dimension. + - sorted_token_ids: A tensor containing the sorted indices of tokens, + repeated topk times and arranged by the expert index they are + assigned to. + - expert_ids: A tensor containing the indices of the expert for each + block. It determines which expert matrix from B should be used for + each block in A. + This kernel performs the multiplication of a token by its corresponding + expert matrix as determined by `expert_ids`. The sorting of + `sorted_token_ids` by expert index and padding ensures divisibility by + BLOCK_SIZE_M, which is necessary to maintain consistency in block matrix + multiplication across different blocks processed by the same expert. + """ + # ----------------------------------------------------------- + # Map program ids `pid` to the block of C it should compute. + # This is done in a grouped ordering to promote L2 data reuse. + pid = tl.program_id(axis=0) + num_pid_m = tl.cdiv(EM, BLOCK_SIZE_M) + num_pid_n = tl.cdiv(N, BLOCK_SIZE_N) + num_pid_in_group = GROUP_SIZE_M * num_pid_n + group_id = pid // num_pid_in_group + first_pid_m = group_id * GROUP_SIZE_M + group_size_m = min(num_pid_m - first_pid_m, GROUP_SIZE_M) + pid_m = first_pid_m + ((pid % num_pid_in_group) % group_size_m) + pid_n = (pid % num_pid_in_group) // group_size_m + + # ---------------------------------------------------------- + # Create pointers for the first blocks of A and B. + # We will advance this pointer as we move in the K direction + # and accumulate + # `a_ptrs` is a block of [BLOCK_SIZE_M, BLOCK_SIZE_K] pointers + # `b_ptrs` is a block of [BLOCK_SIZE_K, BLOCK_SIZE_N] pointers + num_tokens_post_padded = tl.load(num_tokens_post_padded_ptr) + if pid_m * BLOCK_SIZE_M >= num_tokens_post_padded: + return + offs_token_id = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M).to(tl.int64) + offs_token = tl.load(sorted_token_ids_ptr + offs_token_id) + token_mask = offs_token < num_valid_tokens + + off_experts = tl.load(expert_ids_ptr + pid_m).to(tl.int64) + if off_experts == -1: + # ----------------------------------------------------------- + # Write back zeros to the output when the expert is not + # in the current expert parallel rank. + write_zeros_to_output( + c_ptr, + stride_cm, + stride_cn, + pid_n, + N, + offs_token, + token_mask, + BLOCK_SIZE_M, + BLOCK_SIZE_N, + compute_type, + ) + return + + offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N).to(tl.int64)) % N + offs_k = tl.arange(0, BLOCK_SIZE_K) + a_ptrs = a_ptr + ( + offs_token[:, None] // top_k * stride_am + offs_k[None, :] * stride_ak + ) + + if use_int4_w4a16: + b_ptrs = ( + b_ptr + + off_experts * stride_be + + (offs_k[:, None] // 2) * stride_bk + + offs_bn[None, :] * stride_bn + ) + b_shifter = (offs_k[:, None] % 2) * 4 + elif use_int8_w8a16: + b_ptrs = ( + b_ptr + + off_experts * stride_be + + offs_k[:, None] * stride_bk + + offs_bn[None, :] * stride_bn + ) + + if not has_zp and use_int4_w4a16: + b_zp_num = 8 + if not has_zp and use_int8_w8a16: + b_zp_num = 128 + elif has_zp and use_int4_w4a16: + b_zp_shifter = (offs_bn[None, :] % 2) * 4 + + # ----------------------------------------------------------- + # Iterate to compute a block of the C matrix. + # We accumulate into a `[BLOCK_SIZE_M, BLOCK_SIZE_N]` block + # of fp32 values for higher accuracy. + # `accumulator` will be converted back to fp16 after the loop. + accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32) + for k in range(0, tl.cdiv(K, BLOCK_SIZE_K)): + # Load the next block of A and B, generate a mask by checking the + # K dimension. + + if not block_k_diviable: + k_mask = offs_k[:, None] < K - k * BLOCK_SIZE_K + k_other = 0.0 + else: + k_mask = None + k_other = None + + a = tl.load( + a_ptrs, + mask=token_mask[:, None] & (offs_k[None, :] < K - k * BLOCK_SIZE_K), + other=0.0, + ) + b = tl.load(b_ptrs) + if use_int4_w4a16: + b = (b >> b_shifter) & 0xF + + b_scale_ptrs = ( + b_scale_ptr + + off_experts * stride_bse + + offs_bn[None, :] * stride_bsn + + ((offs_k[:, None] + BLOCK_SIZE_K * k) // group_size) * stride_bsk + ) + b_scale = tl.load(b_scale_ptrs, mask=k_mask, other=k_other) + b_scale = b_scale.to(tl.float32) + + if has_zp and use_int4_w4a16: + offs_k_true = (offs_k[:, None] + BLOCK_SIZE_K * k) // group_size + b_zp_ptrs = ( + b_zp_ptr + + off_experts * stride_bze + + (offs_bn[None, :] // 2) * stride_bzn + + offs_k_true * stride_bzk + ) + b_zp = tl.load(b_zp_ptrs, mask=k_mask, other=k_other) + b_zp = (b_zp >> b_zp_shifter) & 0xF + b_zp = b_zp.to(tl.float32) + elif has_zp and use_int8_w8a16: + offs_k_true = (offs_k[:, None] + BLOCK_SIZE_K * k) // group_size + b_zp_ptrs = ( + b_zp_ptr + + off_experts * stride_bze + + offs_bn[None, :] * stride_bzn + + offs_k_true * stride_bzk + ) + b_zp = tl.load(b_zp_ptrs, mask=k_mask, other=k_other) + b_zp = b_zp.to(tl.float32) + + # We accumulate along the K dimension. + if has_zp: + b = ((b.to(tl.float32) - b_zp) * b_scale).to(compute_type) + else: + b = ((b.to(tl.float32) - b_zp_num) * b_scale).to(compute_type) + accumulator = tl.dot(a, b, acc=accumulator) + + # Advance the ptrs to the next K block. + a_ptrs += BLOCK_SIZE_K * stride_ak + if use_int4_w4a16: + b_ptrs += (BLOCK_SIZE_K // 2) * stride_bk + else: + b_ptrs += BLOCK_SIZE_K * stride_bk + + if MUL_ROUTED_WEIGHT: + moe_weight = tl.load(topk_weights_ptr + offs_token, mask=token_mask, other=0) + accumulator = accumulator * moe_weight[:, None] + + accumulator = accumulator.to(compute_type) + # ----------------------------------------------------------- + # Write back the block of the output + offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N) + c_ptrs = c_ptr + stride_cm * offs_token[:, None] + stride_cn * offs_cn[None, :] + c_mask = token_mask[:, None] & (offs_cn[None, :] < N) + tl.store(c_ptrs, accumulator, mask=c_mask) + + +@triton.jit +def fused_moe_kernel( + # Pointers to matrices + a_ptr, + b_ptr, + c_ptr, + a_scale_ptr, + b_scale_ptr, + topk_weights_ptr, + sorted_token_ids_ptr, + expert_ids_ptr, + num_tokens_post_padded_ptr, + # Matrix dimensions + N, + K, + EM, + num_valid_tokens, + # The stride variables represent how much to increase the ptr by when + # moving by 1 element in a particular dimension. E.g. `stride_am` is + # how much to increase `a_ptr` by to get the element one row down + # (A has M rows). + stride_am, + stride_ak, + stride_be, + stride_bk, + stride_bn, + stride_cm, + stride_cn, + stride_asm, + stride_ask, + stride_bse, + stride_bsk, + stride_bsn, + # Block size for block-wise quantization + group_n: tl.constexpr, + group_k: tl.constexpr, + # Meta-parameters + BLOCK_SIZE_M: tl.constexpr, + BLOCK_SIZE_N: tl.constexpr, + BLOCK_SIZE_K: tl.constexpr, + GROUP_SIZE_M: tl.constexpr, + MUL_ROUTED_WEIGHT: tl.constexpr, + top_k: tl.constexpr, + compute_type: tl.constexpr, + use_fp8_w8a8: tl.constexpr, + use_int8_w8a16: tl.constexpr, +): + """ + Implements the fused computation for a Mixture of Experts (MOE) using + token and expert matrices. + + Key Parameters: + - A: The input tensor representing tokens with shape (*, K), where '*' can + be any shape representing batches and K is the feature dimension of + each token. + - B: The stacked MOE weight tensor with shape (E, N, K), where E is + the number of experts, K is the input feature dimension, and N is + the output feature dimension. + - C: The output cache tensor with shape (M, topk, N), where M is the + total number of tokens post padding, topk is the number of times + each token is repeated, and N is the output feature dimension. + - sorted_token_ids: A tensor containing the sorted indices of tokens, + repeated topk times and arranged by the expert index they are + assigned to. + - expert_ids: A tensor containing the indices of the expert for each + block. It determines which expert matrix from B should be used for + each block in A. + This kernel performs the multiplication of a token by its corresponding + expert matrix as determined by `expert_ids`. The sorting of + `sorted_token_ids` by expert index and padding ensures divisibility by + BLOCK_SIZE_M, which is necessary to maintain consistency in block matrix + multiplication across different blocks processed by the same expert. + """ + # ----------------------------------------------------------- + # Map program ids `pid` to the block of C it should compute. + # This is done in a grouped ordering to promote L2 data reuse. + pid = tl.program_id(axis=0) + num_pid_m = tl.cdiv(EM, BLOCK_SIZE_M) + num_pid_n = tl.cdiv(N, BLOCK_SIZE_N) + num_pid_in_group = GROUP_SIZE_M * num_pid_n + group_id = pid // num_pid_in_group + first_pid_m = group_id * GROUP_SIZE_M + group_size_m = min(num_pid_m - first_pid_m, GROUP_SIZE_M) + pid_m = first_pid_m + ((pid % num_pid_in_group) % group_size_m) + pid_n = (pid % num_pid_in_group) // group_size_m + + # ---------------------------------------------------------- + # Create pointers for the first blocks of A and B. + # We will advance this pointer as we move in the K direction + # and accumulate + # `a_ptrs` is a block of [BLOCK_SIZE_M, BLOCK_SIZE_K] pointers + # `b_ptrs` is a block of [BLOCK_SIZE_K, BLOCK_SIZE_N] pointers + num_tokens_post_padded = tl.load(num_tokens_post_padded_ptr) + if pid_m * BLOCK_SIZE_M >= num_tokens_post_padded: + return + offs_token_id = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M).to(tl.int64) + offs_token = tl.load(sorted_token_ids_ptr + offs_token_id) + token_mask = offs_token < num_valid_tokens + + off_experts = tl.load(expert_ids_ptr + pid_m).to(tl.int64) + if off_experts == -1: + # ----------------------------------------------------------- + # Write back zeros to the output when the expert is not + # in the current expert parallel rank. + write_zeros_to_output( + c_ptr, + stride_cm, + stride_cn, + pid_n, + N, + offs_token, + token_mask, + BLOCK_SIZE_M, + BLOCK_SIZE_N, + compute_type, + ) + return + + offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N).to(tl.int64)) % N + offs_k = tl.arange(0, BLOCK_SIZE_K) + a_ptrs = a_ptr + ( + offs_token[:, None] // top_k * stride_am + offs_k[None, :] * stride_ak + ) + + b_ptrs = ( + b_ptr + + off_experts * stride_be + + (offs_k[:, None] * stride_bk + offs_bn[None, :] * stride_bn) + ) + if use_int8_w8a16: + b_scale_ptrs = ( + b_scale_ptr + off_experts * stride_bse + offs_bn[None, :] * stride_bsn + ) + b_scale = tl.load(b_scale_ptrs) + + if use_fp8_w8a8: + if group_k > 0 and group_n > 0: + a_scale_ptrs = a_scale_ptr + (offs_token // top_k) * stride_asm + offs_bsn = offs_bn // group_n + b_scale_ptrs = ( + b_scale_ptr + off_experts * stride_bse + offs_bsn * stride_bsn + ) + else: + a_scale = tl.load(a_scale_ptr) + b_scale = tl.load(b_scale_ptr + off_experts) + + # ----------------------------------------------------------- + # Iterate to compute a block of the C matrix. + # We accumulate into a `[BLOCK_SIZE_M, BLOCK_SIZE_N]` block + # of fp32 values for higher accuracy. + # `accumulator` will be converted back to fp16 after the loop. + accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32) + for k in range(0, tl.cdiv(K, BLOCK_SIZE_K)): + # Load the next block of A and B, generate a mask by checking the + # K dimension. + a = tl.load( + a_ptrs, + mask=token_mask[:, None] & (offs_k[None, :] < K - k * BLOCK_SIZE_K), + other=0.0, + ) + b = tl.load(b_ptrs, mask=offs_k[:, None] < K - k * BLOCK_SIZE_K, other=0.0) + # We accumulate along the K dimension. + if use_int8_w8a16: + accumulator = tl.dot(a, b.to(compute_type), acc=accumulator) + elif use_fp8_w8a8: + if group_k > 0 and group_n > 0: + k_start = k * BLOCK_SIZE_K + offs_ks = k_start // group_k + a_scale = tl.load( + a_scale_ptrs + offs_ks * stride_ask, mask=token_mask, other=0.0 + ) + b_scale = tl.load(b_scale_ptrs + offs_ks * stride_bsk) + + accumulator += tl.dot(a, b) * a_scale[:, None] * b_scale[None, :] + else: + accumulator = tl.dot(a, b, acc=accumulator) + else: + accumulator += tl.dot(a, b) + # Advance the ptrs to the next K block. + a_ptrs += BLOCK_SIZE_K * stride_ak + b_ptrs += BLOCK_SIZE_K * stride_bk + + if MUL_ROUTED_WEIGHT: + moe_weight = tl.load(topk_weights_ptr + offs_token, mask=token_mask, other=0) + accumulator = accumulator * moe_weight[:, None] + if use_int8_w8a16: + accumulator = (accumulator * b_scale).to(compute_type) + elif use_fp8_w8a8: + if group_k > 0 and group_n > 0: + accumulator = accumulator.to(compute_type) + else: + accumulator = (accumulator * a_scale * b_scale).to(compute_type) + else: + accumulator = accumulator.to(compute_type) + # ----------------------------------------------------------- + # Write back the block of the output + offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N) + c_ptrs = c_ptr + stride_cm * offs_token[:, None] + stride_cn * offs_cn[None, :] + c_mask = token_mask[:, None] & (offs_cn[None, :] < N) + tl.store(c_ptrs, accumulator, mask=c_mask) + + +def invoke_fused_moe_kernel( + A: torch.Tensor, + B: torch.Tensor, + C: torch.Tensor, + A_scale: Optional[torch.Tensor], + B_scale: Optional[torch.Tensor], + B_zp: Optional[torch.Tensor], + topk_weights: Optional[torch.Tensor], + sorted_token_ids: torch.Tensor, + expert_ids: torch.Tensor, + num_tokens_post_padded: torch.Tensor, + mul_routed_weight: bool, + top_k: int, + config: Dict[str, Any], + compute_type: tl.dtype, + use_fp8_w8a8: bool, + use_int8_w8a16: bool, + use_int4_w4a16: bool, + block_shape: Optional[List[int]] = None, +) -> None: + assert topk_weights is not None or not mul_routed_weight + assert topk_weights is None or topk_weights.stride(1) == 1 + assert sorted_token_ids.stride(0) == 1 + + if use_fp8_w8a8: + assert B_scale is not None + assert ( + block_shape is None + or triton.cdiv(B.shape[-2], block_shape[0]) == B_scale.shape[-2] + ) + assert ( + block_shape is None + or triton.cdiv(B.shape[-1], block_shape[1]) == B_scale.shape[-1] + ) + + elif use_int8_w8a16 or use_int4_w4a16: + assert B_scale is not None + assert block_shape is None or block_shape[0] == 0 + else: + assert A_scale is None + assert B_scale is None + + M = A.shape[0] + num_tokens = M * top_k + + EM = sorted_token_ids.shape[0] + if A.shape[0] < config["BLOCK_SIZE_M"]: + # optimize for small batch_size. + # We assume that top_ids of each token is unique, so + # so num_valid_experts <= batch_size <= BLOCK_SIZE_M, + # and we can skip some invalid blocks. + EM = min(sorted_token_ids.shape[0], A.shape[0] * top_k * config["BLOCK_SIZE_M"]) + grid = lambda META: ( + triton.cdiv(EM, META["BLOCK_SIZE_M"]) + * triton.cdiv(B.shape[1], META["BLOCK_SIZE_N"]), + ) + + if ( + (use_int8_w8a16 or use_int4_w4a16) + and block_shape is not None + and block_shape[1] > 0 + ): + assert B_scale is not None and B_scale.ndim == 3 + assert B_zp is None or B_zp.ndim == 3 + + use_moe_wna16_cuda = should_moe_wna16_use_cuda( + num_valid_tokens=num_tokens, + group_size=block_shape[1], + num_experts=B.shape[0], + bit=4 if use_int4_w4a16 else 8, + ) + config = config.copy() + config.update( + get_moe_wna16_block_config( + config=config, + use_moe_wna16_cuda=use_moe_wna16_cuda, + num_valid_tokens=num_tokens, + size_k=A.shape[1], + size_n=B.shape[1], + num_experts=B.shape[1], + group_size=block_shape[1], + real_top_k=top_k, + block_size_m=config["BLOCK_SIZE_M"], + ) + ) + + if use_moe_wna16_cuda: + bit = 4 if use_int4_w4a16 else 8 + ops.moe_wna16_gemm( + A, + C, + B, + B_scale, + B_zp, + topk_weights if mul_routed_weight else None, + sorted_token_ids, + expert_ids, + num_tokens_post_padded, + top_k, + config["BLOCK_SIZE_M"], + config["BLOCK_SIZE_N"], + config["BLOCK_SIZE_K"], + bit, + ) + return + + fused_moe_kernel_gptq_awq[grid]( + A, + B, + C, + B_scale, + B_zp, + topk_weights, + sorted_token_ids, + expert_ids, + num_tokens_post_padded, + B.shape[1], + A.shape[1], + EM, + num_tokens, + A.stride(0), + A.stride(1), + B.stride(0), + B.stride(2), + B.stride(1), + C.stride(1), + C.stride(2), + B_scale.stride(0), + B_scale.stride(2), + B_scale.stride(1), + B_zp.stride(0) if B_zp is not None else 0, + B_zp.stride(2) if B_zp is not None else 0, + B_zp.stride(1) if B_zp is not None else 0, + block_k_diviable=A.shape[1] % config["BLOCK_SIZE_K"] == 0, + group_size=block_shape[1], + MUL_ROUTED_WEIGHT=mul_routed_weight, + top_k=top_k, + compute_type=compute_type, + has_zp=B_zp is not None, + use_int4_w4a16=use_int4_w4a16, + use_int8_w8a16=use_int8_w8a16, + **config, + ) + else: + config = config.copy() + BLOCK_SIZE_K = config.pop("BLOCK_SIZE_K") + if block_shape is not None: + BLOCK_SIZE_K = min(BLOCK_SIZE_K, min(block_shape[0], block_shape[1])) + fused_moe_kernel[grid]( + A, + B, + C, + A_scale, + B_scale, + topk_weights, + sorted_token_ids, + expert_ids, + num_tokens_post_padded, + B.shape[1], + B.shape[2], + EM, + num_tokens, + A.stride(0), + A.stride(1), + B.stride(0), + B.stride(2), + B.stride(1), + C.stride(1), + C.stride(2), + A_scale.stride(0) if A_scale is not None and A_scale.ndim == 2 else 0, + A_scale.stride(1) if A_scale is not None and A_scale.ndim == 2 else 0, + B_scale.stride(0) if B_scale is not None and B_scale.ndim >= 2 else 0, + B_scale.stride(2) if B_scale is not None and B_scale.ndim == 3 else 0, + B_scale.stride(1) if B_scale is not None and B_scale.ndim >= 2 else 0, + 0 if block_shape is None else block_shape[0], + 0 if block_shape is None else block_shape[1], + MUL_ROUTED_WEIGHT=mul_routed_weight, + top_k=top_k, + compute_type=compute_type, + use_fp8_w8a8=use_fp8_w8a8, + use_int8_w8a16=use_int8_w8a16, + BLOCK_SIZE_K=BLOCK_SIZE_K, + **config, + ) + + +# Adapted from: https://github.com/sgl-project/sglang/pull/2628 +def get_config_file_name( + E: int, N: int, dtype: Optional[str], block_shape: Optional[List[int]] = None +) -> str: + device_name = current_platform.get_device_name().replace(" ", "_") + dtype_selector = "" if not dtype else f",dtype={dtype}" + block_shape_selector = ( + "" if not block_shape or not all(block_shape) else f",block_shape={block_shape}" + ).replace(" ", "") + return f"E={E},N={N},device_name={device_name}{dtype_selector}{block_shape_selector}.json" # noqa: E501 + + +# Adapted from: https://github.com/sgl-project/sglang/pull/2628 +@functools.lru_cache +def get_moe_configs( + E: int, + N: int, + dtype: Optional[str], + block_n: Optional[int] = None, + block_k: Optional[int] = None, +) -> Optional[Dict[int, Any]]: + """ + Return optimized configurations for the fused MoE kernel. + + The return value will be a dictionary that maps an irregular grid of + batch sizes to configurations of the fused_moe kernel. To evaluate the + kernel on a given batch size bs, the closest batch size in the grid should + be picked and the associated configuration chosen to invoke the kernel. + """ + + # First look up if an optimized configuration is available in the configs + # directory + block_shape = [block_n, block_k] if block_n and block_k else None + json_file_name = get_config_file_name(E, N, dtype, block_shape) + + config_file_path = os.path.join( + os.path.dirname(os.path.realpath(__file__)), "configs", json_file_name + ) + if os.path.exists(config_file_path): + with open(config_file_path) as f: + logger.info("Using configuration from %s for MoE layer.", config_file_path) + # If a configuration has been found, return it + return {int(key): val for key, val in json.load(f).items()} + + # If no optimized configuration is available, we will use the default + # configuration + logger.warning( + ( + "Using default MoE config. Performance might be sub-optimal! " + "Config file not found at %s" + ), + config_file_path, + ) + return None + + +def get_moe_wna16_block_config( + config: Dict[str, int], + use_moe_wna16_cuda: bool, + num_valid_tokens: int, + size_k: int, + size_n: int, + num_experts: int, + group_size: int, + real_top_k: int, + block_size_m: int, +): + if "BLOCK_SIZE_N" in config and "BLOCK_SIZE_K" in config: + # optimal block config is set + return {} + if not use_moe_wna16_cuda: + # triton moe wna16 kernel + if num_valid_tokens // real_top_k == 1: + # if bs=1, use a smaller BLOCK_SIZE_N + return {"BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64} + else: + return {"BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32} + else: + # cuda moe wna16 kernel + # set default block_size 128, and increase them when num_blocks + # is too large. + block_size_n = 128 + block_size_k = 128 + if block_size_k <= group_size: + block_size_k = group_size + + num_n_blocks = size_k // block_size_k + num_k_blocks = size_n // block_size_k + num_m_blocks = ( + num_valid_tokens + block_size_m - 1 + ) / block_size_m + num_experts + if num_valid_tokens // real_top_k <= block_size_m: + num_m_blocks = min(num_m_blocks, num_valid_tokens) + num_blocks = num_m_blocks * num_n_blocks * num_k_blocks + + if size_k % 256 == 0 and num_blocks >= 256 and block_size_k < 256: + block_size_k = 256 + num_blocks = num_blocks // (256 // block_size_k) + + if ( + num_m_blocks <= 16 + and size_k % (block_size_k * 2) == 0 + and size_k % (block_size_k * 2) == 0 + and block_size_k <= 512 + and num_blocks >= 512 + ): + block_size_k = block_size_k * 2 + num_blocks = num_blocks // 2 + + if num_blocks > 1024: + block_size_n = 256 + num_n_blocks = num_n_blocks // 2 + num_blocks = num_blocks // 2 + + if size_n <= 1024 and num_blocks >= 1024: + # The kernel performance got much better with BLOCK_SIZE_N=1024 + # when num_blocks is large, event when N is small. + # Not sure why, maybe it force the CUDA SM process only one block + # at the same time. + block_size_n = 1024 + + return {"BLOCK_SIZE_N": block_size_n, "BLOCK_SIZE_K": block_size_k} + + +def should_moe_wna16_use_cuda( + num_valid_tokens: int, group_size: int, num_experts: int, bit: int +): + return ( + bit == 4 and group_size in [32, 64, 128] and num_valid_tokens / num_experts <= 6 + ) + + +def get_default_config( + M: int, + E: int, + N: int, + K: int, + topk: int, + dtype: Optional[str], + is_marlin: bool, + block_shape: Optional[List[int]] = None, +) -> Dict[str, int]: + if dtype == "fp8_w8a8" and block_shape is not None: + # Block-wise quant: BLOCK_SIZE_N must be divisible by block_shape[0] + # BLOCK_SIZE_K must be divisible by block_shape[1] + config = { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": block_shape[0], + "BLOCK_SIZE_K": block_shape[1], + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3, + } + elif dtype in ["int4_w4a16", "int8_w8a16"] and block_shape is not None: + # moe wna16 kernels + # only set BLOCK_SIZE_M + # BLOCK_SIZE_N and BLOCK_SIZE_K would be set later + bit = 4 if dtype == "int4_w4a16" else 8 + use_moe_wna16_cuda = should_moe_wna16_use_cuda(M * topk, block_shape[1], E, bit) + if use_moe_wna16_cuda: + config = {"BLOCK_SIZE_M": min(16, M)} + elif M <= 20: + config = {"BLOCK_SIZE_M": 16, "GROUP_SIZE_M": 1} + elif M <= 40: + config = {"BLOCK_SIZE_M": 32, "GROUP_SIZE_M": 1} + else: + config = {"BLOCK_SIZE_M": 64, "GROUP_SIZE_M": 1} + else: + config = { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 8, + } + # A heuristic: fused marlin works faster with this config for small M + if M <= E or (is_marlin and M <= 32): + config = { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + } + return config + + +def try_get_optimal_moe_config( + w1_shape: Tuple[int, ...], + w2_shape: Tuple[int, ...], + top_k: int, + dtype: Optional[str], + M: int, + is_marlin: bool = False, + block_shape: Optional[List[int]] = None, +): + # from vllm.model_executor.layers.fused_moe import get_config + # TODO: removed when syncing to vLLM, do we need this? + # override_config = get_config() + override_config = None + if override_config: + config = override_config + else: + # First try to load optimal config from the file + E, _, N = w2_shape + if dtype == "int4_w4a16": + N = N * 2 + block_n = block_shape[0] if block_shape else 0 + block_k = block_shape[1] if block_shape else 0 + configs = get_moe_configs(E, N, dtype, block_n, block_k) + + if configs: + # If an optimal configuration map has been found, look up the + # optimal config + config = configs[min(configs.keys(), key=lambda x: abs(x - M))] + else: + # Else use the default config + config = get_default_config( + M, E, N, w1_shape[2], top_k, dtype, is_marlin, block_shape + ) + return config + + +def vllm_topk_softmax( + topk_weights: torch.Tensor, + topk_indices: torch.Tensor, + token_expert_indices: torch.Tensor, + gating_output: torch.Tensor, + renormalize: bool, +) -> tuple[torch.Tensor, ...]: + ops.topk_softmax( + topk_weights, + topk_indices, + token_expert_indices, + gating_output, + ) + if renormalize: + topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True) + + return topk_weights, topk_indices + + +def dispatch_topk_func() -> Callable[..., tuple[torch.Tensor, ...]]: + return vllm_topk_softmax + + +def fused_topk( + hidden_states: torch.Tensor, + gating_output: torch.Tensor, + topk: int, + renormalize: bool, +) -> Tuple[torch.Tensor, torch.Tensor]: + assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch" + + M, _ = hidden_states.shape + + topk_weights = torch.empty( + M, topk, dtype=torch.float32, device=hidden_states.device + ) + topk_ids = torch.empty(M, topk, dtype=torch.int32, device=hidden_states.device) + token_expert_indicies = torch.empty( + M, topk, dtype=torch.int32, device=hidden_states.device + ) + + gating_output_float = gating_output.float() # TODO(woosuk): Optimize this. + + topk_func = dispatch_topk_func() + topk_weights, topk_ids = topk_func( + topk_weights, topk_ids, token_expert_indicies, gating_output_float, renormalize + ) + + del token_expert_indicies # Not used. Will be used in the future. + return topk_weights, topk_ids + + +# This is used by the Deepseek-V2 and Deepseek-V3 model +@torch.compile(dynamic=True, backend=current_platform.simple_compile_backend) +def grouped_topk( + hidden_states: torch.Tensor, + gating_output: torch.Tensor, + topk: int, + renormalize: bool, + num_expert_group: int = 0, + topk_group: int = 0, + scoring_func: str = "softmax", + e_score_correction_bias: Optional[torch.Tensor] = None, +) -> Tuple[torch.Tensor, torch.Tensor]: + + assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch" + + if scoring_func == "softmax": + scores = torch.softmax(gating_output, dim=-1) + elif scoring_func == "sigmoid": + scores = gating_output.sigmoid() + else: + raise ValueError(f"Unsupported scoring function: {scoring_func}") + + num_token = scores.shape[0] + if e_score_correction_bias is not None: + # Store original scores before applying correction bias. We use biased + # scores for expert selection but original scores for routing weights + original_scores = scores + scores = scores + e_score_correction_bias.unsqueeze(0) + group_scores = ( + scores.view(num_token, num_expert_group, -1).topk(2, dim=-1)[0].sum(dim=-1) + ) + else: + group_scores = ( + scores.view(num_token, num_expert_group, -1).max(dim=-1).values + ) # [n, n_group] + group_idx = torch.topk(group_scores, k=topk_group, dim=-1, sorted=False)[ + 1 + ] # [n, top_k_group] + group_mask = torch.zeros_like(group_scores) # [n, n_group] + group_mask.scatter_(1, group_idx, 1) # [n, n_group] + score_mask = ( + group_mask.unsqueeze(-1) + .expand(num_token, num_expert_group, scores.shape[-1] // num_expert_group) + .reshape(num_token, -1) + ) # [n, e] + tmp_scores = scores.masked_fill(~score_mask.bool(), float("-inf")) # [n, e] + + if e_score_correction_bias is not None: + topk_ids = torch.topk(tmp_scores, k=topk, dim=-1, sorted=False)[1] + # Use original unbiased scores for the routing weights + topk_weights = original_scores.gather(1, topk_ids) + else: + topk_weights, topk_ids = torch.topk(tmp_scores, k=topk, dim=-1, sorted=False) + + if renormalize: + topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True) + + return topk_weights.to(torch.float32), topk_ids.to(torch.int32) + + +def get_config_dtype_str( + dtype: torch.dtype, + use_int4_w4a16: Optional[bool] = False, + use_int8_w8a16: Optional[bool] = False, + use_fp8_w8a8: Optional[bool] = False, +) -> Optional[str]: + if use_fp8_w8a8: + return "fp8_w8a8" + elif use_int8_w8a16: + return "int8_w8a16" + elif use_int4_w4a16: + return "int4_w4a16" + elif dtype == torch.float: + # avoiding cases where kernel fails when float32 MoE + # use fp16/bfloat16 configs + return "float32" + return None + + +def inplace_fused_experts( + hidden_states: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + activation: str = "silu", + apply_router_weight_on_input: bool = False, + use_fp8_w8a8: bool = False, + use_int8_w8a16: bool = False, + use_int4_w4a16: bool = False, + global_num_experts: int = -1, + expert_map: Optional[torch.Tensor] = None, + w1_scale: Optional[torch.Tensor] = None, + w2_scale: Optional[torch.Tensor] = None, + w1_zp: Optional[torch.Tensor] = None, + w2_zp: Optional[torch.Tensor] = None, + a1_scale: Optional[torch.Tensor] = None, + a2_scale: Optional[torch.Tensor] = None, + block_shape: Optional[List[int]] = None, +) -> None: + fused_experts_impl( + hidden_states, + w1, + w2, + topk_weights, + topk_ids, + True, + activation, + apply_router_weight_on_input, + use_fp8_w8a8, + use_int8_w8a16, + use_int4_w4a16, + global_num_experts, + expert_map, + w1_scale, + w2_scale, + w1_zp, + w2_zp, + a1_scale, + a2_scale, + block_shape, + ) + + +def inplace_fused_experts_fake( + hidden_states: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + activation: str = "silu", + apply_router_weight_on_input: bool = False, + use_fp8_w8a8: bool = False, + use_int8_w8a16: bool = False, + use_int4_w4a16: bool = False, + global_num_experts: int = -1, + expert_map: Optional[torch.Tensor] = None, + w1_scale: Optional[torch.Tensor] = None, + w2_scale: Optional[torch.Tensor] = None, + w1_zp: Optional[torch.Tensor] = None, + w2_zp: Optional[torch.Tensor] = None, + a1_scale: Optional[torch.Tensor] = None, + a2_scale: Optional[torch.Tensor] = None, + block_shape: Optional[List[int]] = None, +) -> None: + pass + + +direct_register_custom_op( + op_name="inplace_fused_experts", + op_func=inplace_fused_experts, + mutates_args=["hidden_states"], + fake_impl=inplace_fused_experts_fake, +) + + +def outplace_fused_experts( + hidden_states: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + activation: str = "silu", + apply_router_weight_on_input: bool = False, + use_fp8_w8a8: bool = False, + use_int8_w8a16: bool = False, + use_int4_w4a16: bool = False, + global_num_experts: int = -1, + expert_map: Optional[torch.Tensor] = None, + w1_scale: Optional[torch.Tensor] = None, + w2_scale: Optional[torch.Tensor] = None, + w1_zp: Optional[torch.Tensor] = None, + w2_zp: Optional[torch.Tensor] = None, + a1_scale: Optional[torch.Tensor] = None, + a2_scale: Optional[torch.Tensor] = None, + block_shape: Optional[List[int]] = None, +) -> torch.Tensor: + return fused_experts_impl( + hidden_states, + w1, + w2, + topk_weights, + topk_ids, + False, + activation, + apply_router_weight_on_input, + use_fp8_w8a8, + use_int8_w8a16, + use_int4_w4a16, + global_num_experts, + expert_map, + w1_scale, + w2_scale, + w1_zp, + w2_zp, + a1_scale, + a2_scale, + block_shape, + ) + + +def outplace_fused_experts_fake( + hidden_states: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + activation: str = "silu", + apply_router_weight_on_input: bool = False, + use_fp8_w8a8: bool = False, + use_int8_w8a16: bool = False, + use_int4_w4a16: bool = False, + global_num_experts: int = -1, + expert_map: Optional[torch.Tensor] = None, + w1_scale: Optional[torch.Tensor] = None, + w2_scale: Optional[torch.Tensor] = None, + w1_zp: Optional[torch.Tensor] = None, + w2_zp: Optional[torch.Tensor] = None, + a1_scale: Optional[torch.Tensor] = None, + a2_scale: Optional[torch.Tensor] = None, + block_shape: Optional[List[int]] = None, +) -> torch.Tensor: + return torch.empty_like(hidden_states) + + +direct_register_custom_op( + op_name="outplace_fused_experts", + op_func=outplace_fused_experts, + mutates_args=[], + fake_impl=outplace_fused_experts_fake, +) + + +def torch_vllm_inplace_fused_experts(**kwargs) -> torch.Tensor: + ops.inplace_fused_experts(**kwargs) + hidden_states = kwargs["hidden_states"] + return hidden_states + + +def torch_vllm_outplace_fused_experts(**kwargs) -> torch.Tensor: + return ops.outplace_fused_experts(**kwargs) + + +def dispatch_fused_experts_func(inplace: bool) -> Callable[..., torch.Tensor]: + if inplace: + return torch_vllm_inplace_fused_experts + return torch_vllm_outplace_fused_experts + + +def fused_experts( + hidden_states: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + inplace: bool = False, + activation: str = "silu", + apply_router_weight_on_input: bool = False, + use_fp8_w8a8: bool = False, + use_int8_w8a16: bool = False, + use_int4_w4a16: bool = False, + global_num_experts: int = -1, + expert_map: Optional[torch.Tensor] = None, + w1_scale: Optional[torch.Tensor] = None, + w2_scale: Optional[torch.Tensor] = None, + w1_zp: Optional[torch.Tensor] = None, + w2_zp: Optional[torch.Tensor] = None, + a1_scale: Optional[torch.Tensor] = None, + a2_scale: Optional[torch.Tensor] = None, + block_shape: Optional[List[int]] = None, +) -> torch.Tensor: + return dispatch_fused_experts_func(inplace)( + hidden_states=hidden_states, + w1=w1, + w2=w2, + topk_weights=topk_weights, + topk_ids=topk_ids, + activation=activation, + apply_router_weight_on_input=apply_router_weight_on_input, + use_fp8_w8a8=use_fp8_w8a8, + use_int8_w8a16=use_int8_w8a16, + use_int4_w4a16=use_int4_w4a16, + global_num_experts=global_num_experts, + expert_map=expert_map, + w1_scale=w1_scale, + w2_scale=w2_scale, + w1_zp=w1_zp, + w2_zp=w2_zp, + a1_scale=a1_scale, + a2_scale=a2_scale, + block_shape=block_shape, + ) + + +def fused_experts_impl( + hidden_states: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + inplace: bool = False, + activation: str = "silu", + apply_router_weight_on_input: bool = False, + use_fp8_w8a8: bool = False, + use_int8_w8a16: bool = False, + use_int4_w4a16: bool = False, + global_num_experts: int = -1, + expert_map: Optional[torch.Tensor] = None, + w1_scale: Optional[torch.Tensor] = None, + w2_scale: Optional[torch.Tensor] = None, + w1_zp: Optional[torch.Tensor] = None, + w2_zp: Optional[torch.Tensor] = None, + a1_scale: Optional[torch.Tensor] = None, + a2_scale: Optional[torch.Tensor] = None, + block_shape: Optional[List[int]] = None, +): + # Check constraints. + if use_int4_w4a16: + assert hidden_states.shape[1] // 2 == w1.shape[2], "Hidden size mismatch" + else: + assert hidden_states.shape[1] == w1.shape[2], "Hidden size mismatch" + + assert topk_weights.shape == topk_ids.shape, "topk shape mismatch" + assert hidden_states.is_contiguous(), "Hidden_states must be contiguous" + assert w1.stride(-1) == 1, "Stride of last dimension must be 1" + assert w2.stride(-1) == 1, "Stride of last dimension must be 1" + assert hidden_states.dtype in [torch.float32, torch.float16, torch.bfloat16] + + num_tokens, _ = hidden_states.shape + E, N, _ = w1.shape + K = w2.shape[1] + if global_num_experts == -1: + global_num_experts = E + top_k_num = topk_ids.shape[1] + # We execute the fused_moe kernel in chunks to circumvent this issue: + # https://github.com/vllm-project/vllm/issues/5938 + CHUNK_SIZE = VLLM_FUSED_MOE_CHUNK_SIZE + M = min(num_tokens, CHUNK_SIZE) + config_dtype = get_config_dtype_str( + use_fp8_w8a8=use_fp8_w8a8, + use_int8_w8a16=use_int8_w8a16, + use_int4_w4a16=use_int4_w4a16, + dtype=hidden_states.dtype, + ) + + get_config_func = functools.partial( + try_get_optimal_moe_config, + w1.shape, + w2.shape, + top_k_num, + config_dtype, + block_shape=block_shape, + ) + + config = get_config_func(M) + + # We can reuse the memory between these because by the time we need + # cache3, we're done with cache1 + cache13 = torch.empty( + M * top_k_num * max(N, K), + device=hidden_states.device, + dtype=hidden_states.dtype, + ) + intermediate_cache1 = cache13[: M * top_k_num * N].view(M, top_k_num, N) + intermediate_cache3 = cache13[: M * top_k_num * K].view(M, top_k_num, K) + + # This needs separate memory since it's used concurrently with cache1 + intermediate_cache2 = torch.empty( + (M * top_k_num, N // 2), device=hidden_states.device, dtype=hidden_states.dtype + ) + + if hidden_states.dtype == torch.bfloat16: + compute_type = tl.bfloat16 + elif hidden_states.dtype == torch.float16: + compute_type = tl.float16 + elif hidden_states.dtype == torch.float32: + compute_type = tl.float32 + else: + raise ValueError(f"Unsupported compute_type: {hidden_states.dtype}") + + if inplace: + out_hidden_states = hidden_states + else: + out_hidden_states = torch.empty_like(hidden_states) + + for chunk in range((num_tokens // CHUNK_SIZE) + 1): + begin_chunk_idx, end_chunk_idx = ( + chunk * CHUNK_SIZE, + min((chunk + 1) * CHUNK_SIZE, num_tokens), + ) + curr_hidden_states = hidden_states[begin_chunk_idx:end_chunk_idx] + tokens_in_chunk, _ = curr_hidden_states.shape + + if tokens_in_chunk == 0: + break + + if tokens_in_chunk < CHUNK_SIZE and chunk > 0: + # Adjust the intermediate cache size and config for the last + # chunk. Note that in most cases we only have one chunk + # so the cache size and config are already set correctly and + # do not need to be adjusted. + intermediate_cache1 = intermediate_cache1[:tokens_in_chunk] + intermediate_cache2 = intermediate_cache2[ + : tokens_in_chunk * topk_ids.shape[1] + ] + intermediate_cache3 = intermediate_cache3[:tokens_in_chunk] + config = get_config_func(tokens_in_chunk) + + curr_topk_ids = topk_ids[begin_chunk_idx:end_chunk_idx] + curr_topk_weights = topk_weights[begin_chunk_idx:end_chunk_idx] + + a1q_scale: Optional[torch.Tensor] = None + + if use_fp8_w8a8: + qcurr_hidden_states, a1q_scale = _fp8_quantize( + curr_hidden_states, a1_scale, block_shape + ) + else: + qcurr_hidden_states = curr_hidden_states + a1q_scale = a1_scale + + sorted_token_ids, expert_ids, num_tokens_post_padded = moe_align_block_size( + curr_topk_ids, config["BLOCK_SIZE_M"], global_num_experts, expert_map + ) + + invoke_fused_moe_kernel( + qcurr_hidden_states, + w1, + intermediate_cache1, + a1q_scale, + w1_scale, + w1_zp, + curr_topk_weights, + sorted_token_ids, + expert_ids, + num_tokens_post_padded, + apply_router_weight_on_input, + top_k_num, + config, + compute_type=compute_type, + use_fp8_w8a8=use_fp8_w8a8, + use_int8_w8a16=use_int8_w8a16, + use_int4_w4a16=use_int4_w4a16, + block_shape=block_shape, + ) + + if activation == "silu": + ops.silu_and_mul(intermediate_cache2, intermediate_cache1.view(-1, N)) + elif activation == "gelu": + ops.gelu_and_mul(intermediate_cache2, intermediate_cache1.view(-1, N)) + else: + raise ValueError(f"Unsupported FusedMoe activation: {activation}") + + a2q_scale: Optional[torch.Tensor] = None + + if use_fp8_w8a8: + qintermediate_cache2, a2q_scale = _fp8_quantize( + intermediate_cache2, a2_scale, block_shape + ) + else: + qintermediate_cache2 = intermediate_cache2 + a2q_scale = a2_scale + + invoke_fused_moe_kernel( + qintermediate_cache2, + w2, + intermediate_cache3, + a2q_scale, + w2_scale, + w2_zp, + curr_topk_weights, + sorted_token_ids, + expert_ids, + num_tokens_post_padded, + not apply_router_weight_on_input, + 1, + config, + compute_type=compute_type, + use_fp8_w8a8=use_fp8_w8a8, + use_int8_w8a16=use_int8_w8a16, + use_int4_w4a16=use_int4_w4a16, + block_shape=block_shape, + ) + + ops.moe_sum( + intermediate_cache3.view(*intermediate_cache3.shape), + out_hidden_states[begin_chunk_idx:end_chunk_idx], + ) + + return out_hidden_states + + +def fused_moe( + hidden_states: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + gating_output: torch.Tensor, + topk: int, + renormalize: bool, + inplace: bool = False, + activation: str = "silu", + apply_router_weight_on_input: bool = False, + use_grouped_topk: bool = False, + num_expert_group: Optional[int] = None, + topk_group: Optional[int] = None, + custom_routing_function: Optional[Callable] = None, + use_fp8_w8a8: bool = False, + use_int8_w8a16: bool = False, + use_int4_w4a16: bool = False, + global_num_experts: int = -1, + expert_map: Optional[torch.Tensor] = None, + w1_scale: Optional[torch.Tensor] = None, + w2_scale: Optional[torch.Tensor] = None, + w1_zp: Optional[torch.Tensor] = None, + w2_zp: Optional[torch.Tensor] = None, + a1_scale: Optional[torch.Tensor] = None, + a2_scale: Optional[torch.Tensor] = None, + block_shape: Optional[List[int]] = None, +) -> torch.Tensor: + """ + This function computes a Mixture of Experts (MoE) layer using two sets of + weights, w1 and w2, and top-k gating mechanism. + + Parameters: + - hidden_states (torch.Tensor): The input tensor to the MoE layer. + - w1 (torch.Tensor): The first set of expert weights. + - w2 (torch.Tensor): The second set of expert weights. + - gating_output (torch.Tensor): The output of the gating operation + (before softmax). + - topk (int): The number of top-k experts to select. + - renormalize (bool): If True, renormalize the top-k weights to sum to 1. + - inplace (bool): If True, perform the operation in-place. + Defaults to False. + - activation (str): The activation function to apply after the first + MoE layer. + - num_expert_group: Optional[int]: additional parameter for grouped_topk + - topk_group: Optional[int]: additional parameter for grouped_topk + - use_grouped_topk: If True, use grouped_topk instead of fused_topk + note: Deepseekv2 model uses grouped_topk + - use_fp8_w8a8 (bool): If True, use fp8 arithmetic to compute the inner + products for w1 and w2. Defaults to False. + - use_int8_w8a16 (bool): If True, use matmul of int8 weight and bf16/fp16 + activation to compute the inner products for w1 and w2. + Defaults to False. + - use_int4_w4a16 (bool): If True, use matmul of int4 weight and bf16/fp16 + activation to compute the inner products for w1 and w2. + Defaults to False. + - global_num_experts (int): The total number of experts in the global + expert space. + - expert_map (Optional[torch.Tensor]): A tensor mapping expert indices + from the global expert space to the local expert space of the expert + parallel shard. + - w1_scale (Optional[torch.Tensor]): Optional scale to be used for + w1. + - w2_scale (Optional[torch.Tensor]): Optional scale to be used for + w2. + - a1_scale (Optional[torch.Tensor]): Optional scale to be used for + a1. + - a2_scale (Optional[torch.Tensor]): Optional scale to be used for + a2. + - block_shape: (Optional[List[int]]): Optional block size for block-wise + quantization. + + Returns: + - torch.Tensor: The output tensor after applying the MoE layer. + """ + + if use_grouped_topk: + assert num_expert_group is not None and topk_group is not None + topk_weights, topk_ids = grouped_topk( + hidden_states, + gating_output, + topk, + renormalize, + num_expert_group, + topk_group, + ) + elif custom_routing_function is None: + topk_weights, topk_ids = fused_topk( + hidden_states, gating_output, topk, renormalize + ) + else: + topk_weights, topk_ids = custom_routing_function( + hidden_states, gating_output, topk, renormalize + ) + + return fused_experts( + hidden_states, + w1, + w2, + topk_weights, + topk_ids, + inplace=inplace, + activation=activation, + apply_router_weight_on_input=apply_router_weight_on_input, + use_fp8_w8a8=use_fp8_w8a8, + use_int8_w8a16=use_int8_w8a16, + use_int4_w4a16=use_int4_w4a16, + global_num_experts=global_num_experts, + expert_map=expert_map, + w1_scale=w1_scale, + w2_scale=w2_scale, + w1_zp=w1_zp, + w2_zp=w2_zp, + a1_scale=a1_scale, + a2_scale=a2_scale, + block_shape=block_shape, + ) diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/layers.py b/build/torch27-cxx11-cu126-aarch64-linux/moe/layers.py new file mode 100644 index 0000000000000000000000000000000000000000..1befea3f9e97a6a4aa1679821d30091c3d825bb8 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/layers.py @@ -0,0 +1,64 @@ +import torch +import torch.nn as nn + +from .fused_moe import fused_moe + + +def _llama4_topk( + hidden_states: torch.Tensor, + gating_output: torch.Tensor, + topk: int, + renormalize: bool, +): + assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch" + topk_weights, topk_ids = torch.topk(gating_output, topk, dim=1) + topk_weights = torch.sigmoid(topk_weights.float()).to(hidden_states.dtype) + return topk_weights, topk_ids + + +def _fix_llama4_experts(hidden_states: torch.Tensor, experts: nn.Module): + if experts.gate_up_proj.shape[-1] != hidden_states.shape[-1]: + experts.gate_up_proj = nn.Parameter( + experts.gate_up_proj.transpose(1, 2).contiguous() + ) + experts.down_proj = nn.Parameter(experts.down_proj.transpose(1, 2).contiguous()) + + +class Llama4TextMoe(nn.Module): + has_backward = False + + experts: nn.Module + router: nn.Linear + shared_expert: nn.Module + top_k: int + + def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: + hidden_states = hidden_states.reshape(-1, hidden_states.shape[-1]) + + _fix_llama4_experts(hidden_states, self.experts) + + router_logits = self.router(hidden_states) + + extra_kwargs = {} + use_fp8_w8a8 = False + if hasattr(self.experts, "gate_up_proj_scale"): + use_fp8_w8a8 = True + extra_kwargs["w1_scale"] = self.experts.gate_up_proj_scale + extra_kwargs["w2_scale"] = self.experts.down_proj_scale + + out = fused_moe( + hidden_states, + w1=self.experts.gate_up_proj, + w2=self.experts.down_proj, + gating_output=router_logits, + topk=self.top_k, + renormalize=False, + custom_routing_function=_llama4_topk, + apply_router_weight_on_input=True, + use_fp8_w8a8=use_fp8_w8a8, + **extra_kwargs + ) + + out += self.shared_expert(hidden_states) + + return out, router_logits.t() diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/moe_align_block_size.py b/build/torch27-cxx11-cu126-aarch64-linux/moe/moe_align_block_size.py new file mode 100644 index 0000000000000000000000000000000000000000..a698de5aaa123df5a2d663f252e1d5914e880d09 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/moe_align_block_size.py @@ -0,0 +1,252 @@ +# SPDX-License-Identifier: Apache-2.0 +import os +from typing import Optional, Tuple + +import torch +import triton +import triton.language as tl + +from ._ops import ops + + +def round_up(x: int, y: int) -> int: + return ((x + y - 1) // y) * y + + +VLLM_ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON = bool( + int(os.getenv("VLLM_ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON", "0")) +) + + +def ceil_div(a, b): + return (a + b - 1) // b + + +@triton.jit +def moe_align_block_size_stage1( + topk_ids_ptr, + tokens_cnts_ptr, + num_experts: tl.constexpr, + numel: tl.constexpr, + tokens_per_thread: tl.constexpr, +): + pid = tl.program_id(0) + + start_idx = pid * tokens_per_thread + + off_c = (pid + 1) * num_experts + + for i in range(tokens_per_thread): + if start_idx + i < numel: + idx = tl.load(topk_ids_ptr + start_idx + i) + token_cnt = tl.load(tokens_cnts_ptr + off_c + idx) + tl.store(tokens_cnts_ptr + off_c + idx, token_cnt + 1) + + +@triton.jit +def moe_align_block_size_stage2( + tokens_cnts_ptr, + num_experts: tl.constexpr, +): + pid = tl.program_id(0) + + last_cnt = 0 + for i in range(1, num_experts + 1): + token_cnt = tl.load(tokens_cnts_ptr + i * num_experts + pid) + last_cnt = last_cnt + token_cnt + tl.store(tokens_cnts_ptr + i * num_experts + pid, last_cnt) + + +@triton.jit +def moe_align_block_size_stage3( + total_tokens_post_pad_ptr, + tokens_cnts_ptr, + cumsum_ptr, + num_experts: tl.constexpr, + block_size: tl.constexpr, +): + last_cumsum = 0 + off_cnt = num_experts * num_experts + for i in range(1, num_experts + 1): + token_cnt = tl.load(tokens_cnts_ptr + off_cnt + i - 1) + last_cumsum = last_cumsum + tl.cdiv(token_cnt, block_size) * block_size + tl.store(cumsum_ptr + i, last_cumsum) + tl.store(total_tokens_post_pad_ptr, last_cumsum) + + +@triton.jit +def moe_align_block_size_stage4( + topk_ids_ptr, + sorted_token_ids_ptr, + expert_ids_ptr, + tokens_cnts_ptr, + cumsum_ptr, + num_experts: tl.constexpr, + block_size: tl.constexpr, + numel: tl.constexpr, + tokens_per_thread: tl.constexpr, +): + pid = tl.program_id(0) + start_idx = tl.load(cumsum_ptr + pid) + end_idx = tl.load(cumsum_ptr + pid + 1) + + for i in range(start_idx, end_idx, block_size): + tl.store(expert_ids_ptr + i // block_size, pid) + + start_idx = pid * tokens_per_thread + off_t = pid * num_experts + + for i in range(start_idx, tl.minimum(start_idx + tokens_per_thread, numel)): + expert_id = tl.load(topk_ids_ptr + i) + token_cnt = tl.load(tokens_cnts_ptr + off_t + expert_id) + rank_post_pad = token_cnt + tl.load(cumsum_ptr + expert_id) + tl.store(sorted_token_ids_ptr + rank_post_pad, i) + tl.store(tokens_cnts_ptr + off_t + expert_id, token_cnt + 1) + + +# Triton implementation based on: +# https://github.com/sgl-project/sglang/commit/ba5112ff691d791a9e38c6c71f59324a5fcb49d0 +def moe_align_block_size_triton( + topk_ids: torch.Tensor, + num_experts: int, + block_size: int, + sorted_token_ids: torch.Tensor, + expert_ids: torch.Tensor, + num_tokens_post_pad: torch.Tensor, +) -> None: + numel = topk_ids.numel() + grid = (num_experts,) + tokens_cnts = torch.zeros( + (num_experts + 1, num_experts), dtype=torch.int32, device=topk_ids.device + ) + cumsum = torch.zeros((num_experts + 1,), dtype=torch.int32, device=topk_ids.device) + tokens_per_thread = ceil_div(numel, num_experts) + + moe_align_block_size_stage1[grid]( + topk_ids, + tokens_cnts, + num_experts, + numel, + tokens_per_thread, + ) + moe_align_block_size_stage2[grid]( + tokens_cnts, + num_experts, + ) + moe_align_block_size_stage3[(1,)]( + num_tokens_post_pad, + tokens_cnts, + cumsum, + num_experts, + block_size, + ) + moe_align_block_size_stage4[grid]( + topk_ids, + sorted_token_ids, + expert_ids, + tokens_cnts, + cumsum, + num_experts, + block_size, + numel, + tokens_per_thread, + ) + + +def moe_align_block_size( + topk_ids: torch.Tensor, + block_size: int, + num_experts: int, + expert_map: Optional[torch.Tensor] = None, + pad_sorted_ids: bool = False, +) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + """ + Aligns the token distribution across experts to be compatible with block + size for matrix multiplication. + + Parameters: + - topk_ids: A tensor of shape [total_tokens, top_k] representing the + top-k expert indices for each token. + - block_size: The block size used in block matrix multiplication. + - num_experts: The total number of experts. + - expert_map: A tensor of shape [num_experts] that maps the expert index + from the global space to the local index space of the current + expert parallel shard. If the expert is not in the current expert + parallel shard, the mapping is set to -1. + - pad_sorted_ids: A flag indicating whether the sorted_token_ids length + should be padded to a multiple of block_size, + + Returns: + - sorted_token_ids: A tensor containing the sorted token indices according + to their allocated expert. + - expert_ids: A tensor indicating the assigned expert index for each block. + - num_tokens_post_padded: The total number of tokens after padding, + ensuring divisibility by block_size. + + This function pads the number of tokens that each expert needs to process + so that it is divisible by block_size. + Padding ensures that during block matrix multiplication, the dimensions + align correctly. + + Example: + Given topk_ids = [[2, 3, 4], [1, 2, 4], [1, 3, 4], [1, 2, 3]], + block_size = 4, and num_experts = 4: + - We initially have 12 tokens (after repeating 'top_k' times) and 4 experts, + with each expert needing to process 3 tokens. + - As block_size is 4, we pad 1 token for each expert. + - First, flatten topk_ids to [2, 3, 4, 1, 2, 4, 1, 3, 4, 1, 2, 3]. + - Then append padding tokens [12, 12, 12, 12] for each block. + - After sorting by expert index, we obtain token_ids + [3, 6, 9, 12, 0, 4, 10, 12, 1, 7, 11, 12, 2, 5, 8, 12]. + Tokens 12 are non-existent (padding) and are ignored in + the subsequent matrix multiplication. + - The padding ensures that the total number of tokens is now divisible + by block_size for proper block matrix operations. + """ + max_num_tokens_padded = topk_ids.numel() + num_experts * (block_size - 1) + if pad_sorted_ids: + max_num_tokens_padded = round_up(max_num_tokens_padded, block_size) + sorted_ids = torch.empty( + (max_num_tokens_padded,), dtype=torch.int32, device=topk_ids.device + ) + sorted_ids.fill_(topk_ids.numel()) + max_num_m_blocks = triton.cdiv(max_num_tokens_padded, block_size) + # Expert ids must be zeroed out to prevent index out of bounds error while + # mapping global expert ids to local expert ids in expert parallelism. + expert_ids = torch.zeros( + (max_num_m_blocks,), dtype=torch.int32, device=topk_ids.device + ) + num_tokens_post_pad = torch.empty((1), dtype=torch.int32, device=topk_ids.device) + if num_experts >= 224: + if VLLM_ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON or num_experts != 256: + moe_align_block_size_triton( + topk_ids, + num_experts, + block_size, + sorted_ids, + expert_ids, + num_tokens_post_pad, + ) + else: + # Currently requires num_experts=256 + ops.sgl_moe_align_block_size( + topk_ids, + num_experts, + block_size, + sorted_ids, + expert_ids, + num_tokens_post_pad, + ) + else: + ops.moe_align_block_size( + topk_ids, + num_experts, + block_size, + sorted_ids, + expert_ids, + num_tokens_post_pad, + ) + if expert_map is not None: + expert_ids = expert_map[expert_ids] + + return sorted_ids, expert_ids, num_tokens_post_pad diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/platforms.py b/build/torch27-cxx11-cu126-aarch64-linux/moe/platforms.py new file mode 100644 index 0000000000000000000000000000000000000000..db83d0f2b0fadb15c14e8887ac27e4a07b6c4d62 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/platforms.py @@ -0,0 +1,99 @@ +import random +from abc import ABC, abstractmethod +from functools import lru_cache +from typing import NamedTuple, Optional + +import torch + +IS_ROCM = torch.version.hip is not None + + +class DeviceCapability(NamedTuple): + major: int + minor: int + + def as_version_str(self) -> str: + return f"{self.major}.{self.minor}" + + def to_int(self) -> int: + """ + Express device capability as an integer ````. + + It is assumed that the minor version is always a single digit. + """ + assert 0 <= self.minor < 10 + return self.major * 10 + self.minor + + +class Platform(ABC): + simple_compile_backend: str = "inductor" + + @classmethod + @abstractmethod + def get_device_name(cls, device_id: int = 0) -> str: ... + + @abstractmethod + def is_cuda_alike(self) -> bool: + """Stateless version of :func:`torch.cuda.is_available`.""" + ... + + @abstractmethod + def is_rocm(self): ... + + @classmethod + def seed_everything(cls, seed: Optional[int] = None) -> None: + """ + Set the seed of each random module. + `torch.manual_seed` will set seed on all devices. + + Loosely based on: https://github.com/Lightning-AI/pytorch-lightning/blob/2.4.0/src/lightning/fabric/utilities/seed.py#L20 + """ + if seed is not None: + random.seed(seed) + torch.manual_seed(seed) + try: + import numpy as np + + np.random.seed(seed) + except ImportError: + pass + + +class CudaPlatform(Platform): + @classmethod + def get_device_capability(cls, device_id: int = 0) -> DeviceCapability: + major, minor = torch.cuda.get_device_capability(device_id) + return DeviceCapability(major=major, minor=minor) + + @classmethod + @lru_cache(maxsize=8) + def get_device_name(cls, device_id: int = 0) -> str: + return torch.cuda.get_device_name(0) + + def is_cuda_alike(self) -> bool: + return True + + def is_rocm(self): + return False + + +class RocmPlatform(Platform): + @classmethod + @lru_cache(maxsize=8) + def get_device_capability(cls, device_id: int = 0) -> DeviceCapability: + major, minor = torch.cuda.get_device_capability(device_id) + return DeviceCapability(major=major, minor=minor) + + @classmethod + @lru_cache(maxsize=8) + def get_device_name(cls, device_id: int = 0) -> str: + return torch.cuda.get_device_name(device_id) + + def is_cuda_alike(self) -> bool: + return True + + def is_rocm(self): + return True + + +current_platform = RocmPlatform() if IS_ROCM else CudaPlatform() diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/scalar_type.py b/build/torch27-cxx11-cu126-aarch64-linux/moe/scalar_type.py new file mode 100644 index 0000000000000000000000000000000000000000..ea749fe8247b6846620ccbba30ddf48d914ca4e1 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/scalar_type.py @@ -0,0 +1,336 @@ +import functools +import struct +from dataclasses import dataclass +from enum import Enum +from typing import Optional, Union + + +# Mirrors enum in `core/scalar_type.hpp` +class NanRepr(Enum): + NONE = 0 # nans are not supported + IEEE_754 = 1 # nans are: Exp all 1s, mantissa not all 0s + EXTD_RANGE_MAX_MIN = 2 # nans are: Exp all 1s, mantissa all 1s + + +# This ScalarType class is a parallel implementation of the C++ ScalarType +# class found in csrc/core/scalar_type.hpp. These two classes should be kept +# in sync until the inductor fully supports custom C++ classes. +@dataclass(frozen=True) +class ScalarType: + """ + ScalarType can represent a wide range of floating point and integer + types, in particular it can be used to represent sub-byte data types + (something that torch.dtype currently does not support). It is also + capable of representing types with a bias, i.e.: + `stored_value = value + bias`, + this is useful for quantized types (e.g. standard GPTQ 4bit uses a bias + of 8). The implementation for this class can be found in + csrc/core/scalar_type.hpp, these type signatures should be kept in sync + with that file. + """ + + exponent: int + """ + Number of bits in the exponent if this is a floating point type + (zero if this an integer type) + """ + + mantissa: int + """ + Number of bits in the mantissa if this is a floating point type, + or the number bits representing an integer excluding the sign bit if + this an integer type. + """ + + signed: bool + "If the type is signed (i.e. has a sign bit)" + + bias: int + """ + bias used to encode the values in this scalar type + (value = stored_value - bias, default 0) for example if we store the + type as an unsigned integer with a bias of 128 then the value 0 will be + stored as 128 and -1 will be stored as 127 and 1 will be stored as 129. + """ + + _finite_values_only: bool = False + """ + Private: if infs are supported, used `has_infs()` instead. + """ + + nan_repr: NanRepr = NanRepr.IEEE_754 + """ + How NaNs are represent in this scalar type, returns NanRepr value. + (not applicable for integer types) + """ + + def _floating_point_max_int(self) -> int: + assert ( + self.mantissa <= 52 and self.exponent <= 11 + ), f"Cannot represent max/min as a double for type {self.__str__()}" + + max_mantissa = (1 << self.mantissa) - 1 + if self.nan_repr == NanRepr.EXTD_RANGE_MAX_MIN: + max_mantissa = max_mantissa - 1 + + max_exponent = (1 << self.exponent) - 2 + if self.nan_repr == NanRepr.EXTD_RANGE_MAX_MIN or self.nan_repr == NanRepr.NONE: + assert ( + self.exponent < 11 + ), f"Cannot represent max/min as a double for type {self.__str__()}" + max_exponent = max_exponent + 1 + + # adjust the exponent to match that of a double + # for now we assume the exponent bias is the standard 2^(e-1) -1, (where + # e is the exponent bits), there is some precedent for non-standard + # biases, example `float8_e4m3b11fnuz` here: + # https://github.com/jax-ml/ml_dtypes but to avoid premature over + # complication we are just assuming the standard exponent bias until + # there is a need to support non-standard biases + exponent_bias = (1 << (self.exponent - 1)) - 1 + exponent_bias_double = (1 << 10) - 1 # double e = 11 + + max_exponent_double = max_exponent - exponent_bias + exponent_bias_double + + # shift the mantissa and exponent into the proper positions for an + # IEEE double and bitwise-or them together. + return (max_mantissa << (52 - self.mantissa)) | (max_exponent_double << 52) + + def _floating_point_max(self) -> float: + double_raw = self._floating_point_max_int() + return struct.unpack("!d", struct.pack("!Q", double_raw))[0] + + def _raw_max(self) -> Union[int, float]: + if self.is_floating_point(): + return self._floating_point_max() + else: + assert ( + self.size_bits < 64 or self.size_bits == 64 and self.is_signed() + ), "Cannot represent max as an int" + return (1 << self.mantissa) - 1 + + def _raw_min(self) -> Union[int, float]: + if self.is_floating_point(): + assert ( + self.is_signed() + ), "We currently assume all floating point types are signed" + sign_bit_double = 1 << 63 + + max_raw = self._floating_point_max_int() + min_raw = max_raw | sign_bit_double + return struct.unpack("!d", struct.pack("!Q", min_raw))[0] + else: + assert ( + not self.is_signed() or self.size_bits <= 64 + ), "Cannot represent min as a int64_t" + + if self.is_signed(): + return -(1 << (self.size_bits - 1)) + else: + return 0 + + @functools.cached_property + def id(self) -> int: + """ + Convert the ScalarType to an int which can be passed to pytorch custom + ops. This layout of the int must be kept in sync with the C++ + ScalarType's from_id method. + """ + val = 0 + offset = 0 + + def or_and_advance(member, bit_width): + nonlocal val + nonlocal offset + bit_mask = (1 << bit_width) - 1 + val = val | (int(member) & bit_mask) << offset + offset = offset + bit_width + + or_and_advance(self.exponent, 8) + or_and_advance(self.mantissa, 8) + or_and_advance(self.signed, 1) + or_and_advance(self.bias, 32) + or_and_advance(self._finite_values_only, 1) + or_and_advance(self.nan_repr.value, 8) + + assert offset <= 64, f"ScalarType fields too big {offset} to fit into an int64" + + return val + + @property + def size_bits(self) -> int: + return self.exponent + self.mantissa + int(self.signed) + + def min(self) -> Union[int, float]: + """ + Min representable value for this scalar type. + (accounting for bias if there is one) + """ + return self._raw_min() - self.bias + + def max(self) -> Union[int, float]: + """ + Max representable value for this scalar type. + (accounting for bias if there is one) + """ + return self._raw_max() - self.bias + + def is_signed(self) -> bool: + """ + If the type is signed (i.e. has a sign bit), same as `signed` + added for consistency with: + https://pytorch.org/docs/stable/generated/torch.Tensor.is_signed.html + """ + return self.signed + + def is_floating_point(self) -> bool: + "If the type is a floating point type" + return self.exponent != 0 + + def is_integer(self) -> bool: + "If the type is an integer type" + return self.exponent == 0 + + def has_bias(self) -> bool: + "If the type has a non-zero bias" + return self.bias != 0 + + def has_infs(self) -> bool: + "If the type is floating point and supports infinity" + return not self._finite_values_only + + def has_nans(self) -> bool: + return self.nan_repr != NanRepr.NONE.value + + def is_ieee_754(self) -> bool: + """ + If the type is a floating point type that follows IEEE 754 + conventions + """ + return self.nan_repr == NanRepr.IEEE_754.value and not self._finite_values_only + + def __str__(self) -> str: + """ + naming generally follows: https://github.com/jax-ml/ml_dtypes + for floating point types (leading f) the scheme is: + `float_em[flags]` + flags: + - no-flags: means it follows IEEE 754 conventions + - f: means finite values only (no infinities) + - n: means nans are supported (non-standard encoding) + for integer types the scheme is: + `[u]int[b]` + - if bias is not present it means its zero + """ + if self.is_floating_point(): + ret = ( + "float" + + str(self.size_bits) + + "_e" + + str(self.exponent) + + "m" + + str(self.mantissa) + ) + + if not self.is_ieee_754(): + if self._finite_values_only: + ret = ret + "f" + if self.nan_repr != NanRepr.NONE: + ret = ret + "n" + + return ret + else: + ret = ("int" if self.is_signed() else "uint") + str(self.size_bits) + if self.has_bias(): + ret = ret + "b" + str(self.bias) + return ret + + def __repr__(self) -> str: + return "ScalarType." + self.__str__() + + # __len__ needs to be defined (and has to throw TypeError) for pytorch's + # opcheck to work. + def __len__(self) -> int: + raise TypeError + + # + # Convenience Constructors + # + + @classmethod + def int_(cls, size_bits: int, bias: Optional[int]) -> "ScalarType": + "Create a signed integer scalar type (size_bits includes sign-bit)." + ret = cls(0, size_bits - 1, True, bias if bias else 0) + ret.id # noqa B018: make sure the id is cached + return ret + + @classmethod + def uint(cls, size_bits: int, bias: Optional[int]) -> "ScalarType": + """Create a unsigned integer scalar type.""" + ret = cls(0, size_bits, False, bias if bias else 0) + ret.id # noqa B018: make sure the id is cached + return ret + + @classmethod + def float_IEEE754(cls, exponent: int, mantissa: int) -> "ScalarType": + """ + Create a standard floating point type + (i.e. follows IEEE 754 conventions). + """ + assert mantissa > 0 and exponent > 0 + ret = cls(exponent, mantissa, True, 0) + ret.id # noqa B018: make sure the id is cached + return ret + + @classmethod + def float_( + cls, exponent: int, mantissa: int, finite_values_only: bool, nan_repr: NanRepr + ) -> "ScalarType": + """ + Create a non-standard floating point type + (i.e. does not follow IEEE 754 conventions). + """ + assert mantissa > 0 and exponent > 0 + assert nan_repr != NanRepr.IEEE_754, ( + "use `float_IEEE754` constructor for floating point types that " + "follow IEEE 754 conventions" + ) + ret = cls(exponent, mantissa, True, 0, finite_values_only, nan_repr) + ret.id # noqa B018: make sure the id is cached + return ret + + +# naming generally follows: https://github.com/jax-ml/ml_dtypes +# for floating point types (leading f) the scheme is: +# `float_em[flags]` +# flags: +# - no-flags: means it follows IEEE 754 conventions +# - f: means finite values only (no infinities) +# - n: means nans are supported (non-standard encoding) +# for integer types the scheme is: +# `[u]int[b]` +# - if bias is not present it means its zero + + +class scalar_types: + int4 = ScalarType.int_(4, None) + uint4 = ScalarType.uint(4, None) + int8 = ScalarType.int_(8, None) + uint8 = ScalarType.uint(8, None) + float8_e4m3fn = ScalarType.float_(4, 3, True, NanRepr.EXTD_RANGE_MAX_MIN) + float8_e5m2 = ScalarType.float_IEEE754(5, 2) + float16_e8m7 = ScalarType.float_IEEE754(8, 7) + float16_e5m10 = ScalarType.float_IEEE754(5, 10) + + # fp6, https://github.com/usyd-fsalab/fp6_llm/tree/main + float6_e3m2f = ScalarType.float_(3, 2, True, NanRepr.NONE) + + # "gptq" types + uint2b2 = ScalarType.uint(2, 2) + uint3b4 = ScalarType.uint(3, 4) + uint4b8 = ScalarType.uint(4, 8) + uint8b128 = ScalarType.uint(8, 128) + + # colloquial names + bfloat16 = float16_e8m7 + float16 = float16_e5m10 diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/utils/__init__.py b/build/torch27-cxx11-cu126-aarch64-linux/moe/utils/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/utils/custom_ops.py b/build/torch27-cxx11-cu126-aarch64-linux/moe/utils/custom_ops.py new file mode 100644 index 0000000000000000000000000000000000000000..26ef7c23d1bea7a162d81746ac041baefb472f46 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/utils/custom_ops.py @@ -0,0 +1,68 @@ +from typing import Callable, Optional + +import torch +from torch.library import Library + +from .._ops import ops +from ..platforms import current_platform + + +# Some backends use pytorch version < 2.4.0 which doesn't +# support `torch.library.custom_op`. +def supports_custom_op() -> bool: + return hasattr(torch.library, "custom_op") + + +# TODO: we probably need a shorthand for the ops name. +vllm_lib = Library(ops.__name__.split(".")[-1], "FRAGMENT") + + +def direct_register_custom_op( + op_name: str, + op_func: Callable, + mutates_args: list[str], + fake_impl: Optional[Callable] = None, + target_lib: Optional[Library] = None, + dispatch_key: str = "CUDA", +): + """ + `torch.library.custom_op` can have significant overhead because it + needs to consider complicated dispatching logic. This function + directly registers a custom op and dispatches it to the CUDA backend. + See https://gist.github.com/youkaichao/ecbea9ec9fc79a45d2adce1784d7a9a5 + for more details. + + By default, the custom op is registered to the vLLM library. If you + want to register it to a different library, you can pass the library + object to the `target_lib` argument. + + IMPORTANT: the lifetime of the operator is tied to the lifetime of the + library object. If you want to bind the operator to a different library, + make sure the library object is alive when the operator is used. + """ + if not supports_custom_op(): + from ..platforms import current_platform + + assert not current_platform.is_cuda_alike(), ( + "cuda platform needs torch>=2.4 to support custom op, " + "chances are you are using an old version of pytorch " + "or a custom build of pytorch. It is recommended to " + "use vLLM in a fresh new environment and let it install " + "the required dependencies." + ) + return + + import torch.library + + if hasattr(torch.library, "infer_schema"): + schema_str = torch.library.infer_schema(op_func, mutates_args=mutates_args) + else: + # for pytorch 2.4 + import torch._custom_op.impl + + schema_str = torch._custom_op.impl.infer_schema(op_func, mutates_args) + my_lib = target_lib or vllm_lib + my_lib.define(op_name + schema_str) + my_lib.impl(op_name, op_func, dispatch_key=dispatch_key) + if fake_impl is not None: + my_lib._register_fake(op_name, fake_impl) diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/utils/marlin_utils.py b/build/torch27-cxx11-cu126-aarch64-linux/moe/utils/marlin_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..f27b71af67a6f5f18c8dba271926e3d40d518879 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/utils/marlin_utils.py @@ -0,0 +1,306 @@ +from typing import List, Optional, Tuple + +import numpy +import torch + +from ..scalar_type import ScalarType, scalar_types +from .quant_utils import pack_cols, unpack_cols + +GPTQ_MARLIN_TILE = 16 +GPTQ_MARLIN_MIN_THREAD_N = 64 +GPTQ_MARLIN_MIN_THREAD_K = 128 +GPTQ_MARLIN_MAX_PARALLEL = 16 + +GPTQ_MARLIN_24_TILE = 16 +GPTQ_MARLIN_24_MIN_THREAD_N = 128 +GPTQ_MARLIN_24_MIN_THREAD_K = 128 +GPTQ_MARLIN_24_MAX_PARALLEL = 64 + +GPTQ_MARLIN_24_SUPPORTED_QUANT_TYPES = [scalar_types.uint4b8, scalar_types.uint8b128] +GPTQ_MARLIN_24_SUPPORTED_GROUP_SIZES = [-1, 128] + +MARLIN_QQQ_TILE = 16 +MARLIN_QQQ_MIN_THREAD_N = 64 +MARLIN_QQQ_MIN_THREAD_K = 128 +MARLIN_QQQ_MAX_PARALLEL = 16 + +MARLIN_QQQ_SUPPORTED_NUM_BITS = [4] +MARLIN_QQQ_SUPPORTED_GROUP_SIZES = [-1, 128] +MARLIN_QQQ_SUPPORTED_SYM = [True] + +MARLIN_SUPPORTED_GROUP_SIZES = [-1, 32, 64, 128] + +# In case there is a performance issue with Marlin, the variable below can be +# changed to False, which allows Marlin to perform global reductions in fp16 +# precision (instead of fp32), and therefore, save on some memory movements. +USE_FP32_REDUCE_DEFAULT = True + + +# For binary size and compile time, we don't support the same types for with and +# without runtime zero-point. We support common cases, i.e. AWQ and GPTQ. +# TODO: we may want to move this into the C++ so its closer to the actual impl +def query_marlin_supported_quant_types( + has_zp: bool, device_capability: Optional[int] = None +): + if device_capability is None: + capability_tuple = torch.cuda.get_device_capability() + device_capability = capability_tuple[0] * 10 + capability_tuple[1] + + if device_capability < 80: + return [] + + if has_zp: + # AWQ style, unsigned + runtime zero-point + return [scalar_types.uint4, scalar_types.uint8] + else: + # GPTQ style, unsigned + symmetric bias + # TODO: once fp8_marlin is merged into "gptq_marlin" we should be able + # to add `scalar_types.float8_e4m3fn` here + return [scalar_types.uint4b8, scalar_types.uint8b128] + + +def _check_marlin_supported( + quant_type: ScalarType, + group_size: Optional[int], + has_zp: bool, + device_capability: Optional[int] = None, +) -> Tuple[bool, Optional[str]]: + + if device_capability is None: + capability_tuple = torch.cuda.get_device_capability() + device_capability = capability_tuple[0] * 10 + capability_tuple[1] + + supported_types = query_marlin_supported_quant_types(has_zp, device_capability) + + if quant_type not in supported_types: + return ( + False, + f"Marlin does not support weight_bits = {quant_type}. " + f"Only types = {supported_types} " + f"are supported (for group_size = {group_size}, " + f"device_capability = {device_capability}, zp = {has_zp}).", + ) + if group_size is None or group_size not in MARLIN_SUPPORTED_GROUP_SIZES: + return ( + False, + f"Marlin does not support group_size = {group_size}. " + f"Only group_sizes = {MARLIN_SUPPORTED_GROUP_SIZES} " + "are supported.", + ) + + return True, None + + +def check_marlin_supported( + quant_type: ScalarType, + group_size: int, + has_zp: bool = False, + device_capability: Optional[int] = None, +) -> bool: + cond, _ = _check_marlin_supported(quant_type, group_size, has_zp, device_capability) + return cond + + +def verify_marlin_supported( + quant_type: ScalarType, group_size: int, has_zp: bool = False +) -> None: + cond, err_msg = _check_marlin_supported(quant_type, group_size, has_zp) + if not cond: + assert err_msg is not None + raise ValueError(err_msg) + + +def verify_marlin_supports_shape( + output_size_per_partition: int, + input_size_per_partition: int, + input_size: int, + group_size: int, +) -> None: + + # Validate output_size_per_partition + if output_size_per_partition % GPTQ_MARLIN_MIN_THREAD_N != 0: + raise ValueError( + f"Weight output_size_per_partition = " + f"{output_size_per_partition} is not divisible by " + f" min_thread_n = {GPTQ_MARLIN_MIN_THREAD_N}. " + "Consider reducing tensor_parallel_size or running " + "with --quantization gptq." + ) + + # Validate input_size_per_partition + if input_size_per_partition % GPTQ_MARLIN_MIN_THREAD_K != 0: + raise ValueError( + f"Weight input_size_per_partition = " + f"{input_size_per_partition} is not divisible " + f"by min_thread_k = {GPTQ_MARLIN_MIN_THREAD_K}. " + "Consider reducing tensor_parallel_size or running " + "with --quantization gptq." + ) + + if group_size < input_size and input_size_per_partition % group_size != 0: + raise ValueError( + f"Weight input_size_per_partition = {input_size_per_partition}" + f" is not divisible by group_size = {group_size}." + "Consider reducing tensor_parallel_size or running " + "with --quantization gptq." + ) + + +def check_marlin_supports_shape( + output_size_per_partition: int, + input_size_per_partition: int, + input_size: int, + group_size: int, +) -> Tuple[bool, Optional[str]]: + try: + verify_marlin_supports_shape( + output_size_per_partition, input_size_per_partition, input_size, group_size + ) + except ValueError as e: + return False, e.__str__() + return True, None + + +def marlin_make_workspace( + output_size_per_partition: int, device: torch.device +) -> torch.Tensor: + max_workspace_size = ( + output_size_per_partition // GPTQ_MARLIN_MIN_THREAD_N + ) * GPTQ_MARLIN_MAX_PARALLEL + + return torch.zeros( + max_workspace_size, dtype=torch.int, device=device, requires_grad=False + ) + + +def marlin_is_k_full(act_order: bool, is_row_parallel: bool) -> bool: + return (not act_order) or (act_order and not is_row_parallel) + + +def marlin_repeat_scales_on_all_ranks( + act_order: bool, group_size: int, is_row_parallel: bool +) -> bool: + # Need to repeat scales on every rank if act_ordering or + # channelwise and RowParallelLinear + is_channelwise = group_size == -1 + return act_order or (is_channelwise and is_row_parallel) + + +def marlin_make_empty_g_idx(device: torch.device) -> torch.Tensor: + return torch.nn.Parameter( + torch.empty(0, dtype=torch.int, device=device), requires_grad=False + ) + + +def marlin_make_empty_zp(device: torch.device) -> torch.Tensor: + return torch.nn.Parameter( + torch.empty(0, dtype=torch.int, device=device), requires_grad=False + ) + + +def marlin_sort_g_idx(g_idx: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]: + g_idx_sort_indices = torch.argsort(g_idx).to(torch.int) + return g_idx[g_idx_sort_indices], g_idx_sort_indices + + +def get_scale_perms(): + scale_perm: List[int] = [] + for i in range(8): + scale_perm.extend([i + 8 * j for j in range(8)]) + scale_perm_single: List[int] = [] + for i in range(4): + scale_perm_single.extend([2 * i + j for j in [0, 1, 8, 9, 16, 17, 24, 25]]) + return scale_perm, scale_perm_single + + +def marlin_permute_scales( + s: torch.Tensor, size_k: int, size_n: int, group_size: int +) -> torch.Tensor: + + scale_perm, scale_perm_single = get_scale_perms() + if group_size < size_k and group_size != -1: + s = s.reshape((-1, len(scale_perm)))[:, scale_perm] + else: + s = s.reshape((-1, len(scale_perm_single)))[:, scale_perm_single] + s = s.reshape((-1, size_n)).contiguous() + + return s + + +def marlin_moe_permute_scales( + s: torch.Tensor, + size_k: int, + size_n: int, + group_size: int, +): + num_experts = s.shape[0] + output = torch.empty( + (num_experts, s.shape[1], s.shape[2]), + device=s.device, + dtype=s.dtype, + ) + + for e in range(num_experts): + output[e] = marlin_permute_scales(s[e], size_k, size_n, group_size) + return output + + +def marlin_zero_points( + zp: torch.Tensor, size_k: int, size_n: int, num_bits: int +) -> torch.Tensor: + # Permute zero-points in a similar way to scales, but do not use the + # "single" permutation, since zero-points are applied on every MMA + scale_perm, _ = get_scale_perms() + zp = zp.reshape((-1, len(scale_perm)))[:, scale_perm] + + # Interleave column dim (for the dequantize code) and pack it to int32 + if num_bits == 4: + interleave = numpy.array([0, 2, 4, 6, 1, 3, 5, 7]) + elif num_bits == 8: + interleave = numpy.array([0, 2, 1, 3]) + else: + raise Exception("num_bits must be 4 or 8, got {}".format(num_bits)) + + zp = zp.reshape((-1, len(interleave)))[:, interleave].ravel() + zp = zp.reshape((-1, size_n)).contiguous() + zp = pack_cols(zp, num_bits, size_k, size_n) + + return zp + + +def awq_to_marlin_zero_points( + q_zp_packed: torch.Tensor, size_k: int, size_n: int, num_bits: int +) -> torch.Tensor: + # AWQ zero-points are quantized and packed on the column dim. + # In addition, the values are permuted based on dequantizer. + # Here we undo both of these, and then apply marlin permutation + # and pack it back. + q_zp = unpack_cols(q_zp_packed, num_bits, size_k, size_n) + + # Undo interleaving (use argsort(..) to get inverse perm) + if num_bits == 4: + undo_interleave = numpy.argsort(numpy.array([0, 2, 4, 6, 1, 3, 5, 7])) + elif num_bits == 8: + undo_interleave = numpy.argsort(numpy.array([0, 2, 1, 3])) + else: + raise Exception("num_bits must be 4 or 8, got {}".format(num_bits)) + + q_zp = q_zp.reshape((-1, len(undo_interleave)))[:, undo_interleave].ravel() + q_zp = q_zp.reshape((-1, size_n)).contiguous() + + marlin_zp = marlin_zero_points(q_zp, size_k, size_n, num_bits) + return marlin_zp + + +def moe_awq_to_marlin_zero_points( + q_zp_packed: torch.Tensor, size_k: int, size_n: int, num_bits: int +): + num_experts = q_zp_packed.shape[0] + output = torch.empty( + (num_experts, q_zp_packed.shape[1], q_zp_packed.shape[2]), + device=q_zp_packed.device, + dtype=q_zp_packed.dtype, + ) + for e in range(num_experts): + output[e] = awq_to_marlin_zero_points(q_zp_packed[e], size_k, size_n, num_bits) + return output diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/utils/marlin_utils_test.py b/build/torch27-cxx11-cu126-aarch64-linux/moe/utils/marlin_utils_test.py new file mode 100644 index 0000000000000000000000000000000000000000..b03555f5bffa8d48187d37e53f17e6e3138472cc --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/utils/marlin_utils_test.py @@ -0,0 +1,161 @@ +"""Utility functions used for tests and benchmarks""" + +from typing import List, Optional + +import numpy as np +import torch + +from ..scalar_type import ScalarType +from .marlin_utils import GPTQ_MARLIN_TILE, marlin_permute_scales, marlin_zero_points +from .quant_utils import ( + get_pack_factor, + gptq_quantize_weights, + quantize_weights, + sort_weights, +) + + +class MarlinWorkspace: + + def __init__(self, out_features, min_thread_n, max_parallel): + assert ( + out_features % min_thread_n == 0 + ), "out_features = {} is undivisible by min_thread_n = {}".format( + out_features, min_thread_n + ) + + max_workspace_size = (out_features // min_thread_n) * max_parallel + + self.scratch = torch.zeros(max_workspace_size, dtype=torch.int, device="cuda") + + +def marlin_permute_weights(q_w, size_k, size_n, perm, tile=GPTQ_MARLIN_TILE): + assert q_w.shape == (size_k, size_n) + assert size_k % tile == 0, f"size_k = {size_k}, tile = {tile}" + assert size_n % tile == 0, f"size_k = {size_n}, tile = {tile}" + + # Permute weights to 16x64 marlin tiles + q_w = q_w.reshape((size_k // tile, tile, size_n // tile, tile)) + q_w = q_w.permute((0, 2, 1, 3)) + q_w = q_w.reshape((size_k // tile, size_n * tile)) + + q_w = q_w.reshape((-1, perm.numel()))[:, perm].reshape(q_w.shape) + + return q_w + + +def marlin_weights(q_w, size_k, size_n, num_bits, perm): + # Permute + q_w = marlin_permute_weights(q_w, size_k, size_n, perm) + + # Pack + pack_factor = get_pack_factor(num_bits) + orig_device = q_w.device + + q_w = q_w.cpu().numpy().astype(np.uint32) + + q_packed = np.zeros((q_w.shape[0], q_w.shape[1] // pack_factor), dtype=np.uint32) + for i in range(pack_factor): + q_packed |= q_w[:, i::pack_factor] << num_bits * i + + q_packed = torch.from_numpy(q_packed.astype(np.int32)).to(orig_device) + + return q_packed + + +def get_weight_perm(num_bits: int): + perm_list: List[int] = [] + for i in range(32): + perm1: List[int] = [] + col = i // 4 + for block in [0, 1]: + for row in [ + 2 * (i % 4), + 2 * (i % 4) + 1, + 2 * (i % 4 + 4), + 2 * (i % 4 + 4) + 1, + ]: + perm1.append(16 * row + col + 8 * block) + for j in range(4): + perm_list.extend([p + 256 * j for p in perm1]) + + perm = np.array(perm_list) + + if num_bits == 4: + interleave = np.array([0, 2, 4, 6, 1, 3, 5, 7]) + elif num_bits == 8: + interleave = np.array([0, 2, 1, 3]) + else: + raise Exception("num_bits must be 4 or 8, got {}".format(num_bits)) + + perm = perm.reshape((-1, len(interleave)))[:, interleave].ravel() + perm = torch.from_numpy(perm) + return perm + + +def marlin_quantize( + w: torch.Tensor, + quant_type: ScalarType, + group_size: int, + act_order: bool, + test_perm: Optional[torch.Tensor] = None, +): + size_k, size_n = w.shape + num_bits = quant_type.size_bits + + # Normalize group_size + if group_size == -1: + group_size = size_k + assert group_size <= size_k + + # Quantize (and apply act_order if provided) + w_ref, q_w, s, g_idx, rand_perm = gptq_quantize_weights( + w, quant_type, group_size, act_order, test_perm + ) + + # For act_order, sort the "weights" and "g_idx" so that group ids are + # increasing + sort_indices = torch.empty(0, dtype=torch.int, device=w.device) + if act_order: + q_w, g_idx, sort_indices = sort_weights(q_w, g_idx) + + # Reformat to marlin + weight_perm = get_weight_perm(num_bits) + marlin_q_w = marlin_weights(q_w, size_k, size_n, num_bits, weight_perm) + marlin_s = marlin_permute_scales(s, size_k, size_n, group_size) + + # Create result + res_list = [w_ref, marlin_q_w, marlin_s, g_idx, sort_indices, rand_perm] + for i in range(len(res_list)): + res_list[i] = res_list[i].to(w.device) + + return res_list + + +def awq_marlin_quantize(w: torch.Tensor, quant_type: ScalarType, group_size: int): + size_k, size_n = w.shape + + # Normalize group_size + if group_size == -1: + group_size = size_k + assert group_size <= size_k + + # Detect num groups + assert size_k % group_size == 0 + num_groups = size_k // group_size + + # Quantize with zp + w_ref, q_w, s, zp = quantize_weights(w, quant_type, group_size, zero_points=True) + + # Reformat to marlin + weight_perm = get_weight_perm(quant_type.size_bits) + marlin_q_w = marlin_weights(q_w, size_k, size_n, quant_type.size_bits, weight_perm) + marlin_s = marlin_permute_scales(s, size_k, size_n, group_size) + marlin_zp = marlin_zero_points(zp, num_groups, size_n, quant_type.size_bits) + + # Create result + res_list = [w_ref, marlin_q_w, marlin_s, marlin_zp] + for i in range(len(res_list)): + res_list[i] = res_list[i].to(w.device) + + return res_list diff --git a/build/torch27-cxx11-cu126-aarch64-linux/moe/utils/quant_utils.py b/build/torch27-cxx11-cu126-aarch64-linux/moe/utils/quant_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..65c90821773b60cca2b0f9102ca7ad4e3cd009d5 --- /dev/null +++ b/build/torch27-cxx11-cu126-aarch64-linux/moe/utils/quant_utils.py @@ -0,0 +1,470 @@ +"""This file is used for /tests and /benchmarks""" + +from typing import List, Optional + +import numpy +import torch + +from ..scalar_type import ScalarType, scalar_types + +SUPPORTED_GPTQ_QUANT_TYPES = [scalar_types.uint4b8, scalar_types.uint8b128] +SUPPORTED_GROUP_SIZES = [-1, 32, 64, 128] + +MARLIN_QQQ_SUPPORTED_NUM_BITS = [4] + +# Note: this is a hack. We should update each model to register the +# stacked params and get it from there instead in a future PR. +# fused_name: List[shard_name] +FUSED_LAYER_NAME_MAPPING = { + "qkv_proj": ["q_proj", "k_proj", "v_proj"], + "gate_up_proj": ["gate_proj", "up_proj"], +} + + +def pack_quantized_values_into_int32( + w_q: torch.Tensor, wtype: ScalarType, packed_dim: int = 0 +): + # move dim to pack to the end + perm = (*[i for i in range(len(w_q.shape)) if i != packed_dim], packed_dim) + inv_perm = tuple(perm.index(i) for i in range(len(perm))) + w_q_perm = w_q.permute(perm) + + pack_factor = 32 // wtype.size_bits + mask = (1 << wtype.size_bits) - 1 + + new_shape_perm = list(w_q_perm.shape) + assert w_q_perm.shape[-1] % pack_factor == 0 + new_shape_perm[-1] //= pack_factor + + res = torch.zeros(new_shape_perm, dtype=torch.int32, device=w_q.device) + for i in range(pack_factor): + res |= (w_q_perm[..., i::pack_factor] & mask) << wtype.size_bits * i + + return res.permute(inv_perm) + + +def unpack_quantized_values_into_int32( + w_q: torch.Tensor, wtype: ScalarType, packed_dim: int = 0 +): + # move dim to pack to the end + perm = (*[i for i in range(len(w_q.shape)) if i != packed_dim], packed_dim) + inv_perm = tuple(perm.index(i) for i in range(len(perm))) + w_q_perm = w_q.permute(perm) + + pack_factor = 32 // wtype.size_bits + mask = (1 << wtype.size_bits) - 1 + + new_shape_perm = list(w_q_perm.shape) + new_shape_perm[-1] *= pack_factor + + res = torch.zeros(new_shape_perm, dtype=torch.int32, device=w_q.device) + for i in range(pack_factor): + res[..., i::pack_factor] = (w_q_perm >> wtype.size_bits * i) & mask + + return res.permute(inv_perm) + + +def is_layer_skipped(prefix: str, ignored_layers: List[str]) -> bool: + # prefix: model.layers.0.self_attn.q_proj + # proj_name: q_proj + proj_name = prefix.split(".")[-1] + if proj_name in FUSED_LAYER_NAME_MAPPING: + shard_prefixes = [ + prefix.replace(proj_name, shard_proj_name) + for shard_proj_name in FUSED_LAYER_NAME_MAPPING[proj_name] + ] + + is_skipped = None + for shard_prefix in shard_prefixes: + is_shard_skipped = shard_prefix in ignored_layers + + if is_skipped is None: + is_skipped = is_shard_skipped + elif is_shard_skipped != is_skipped: + raise ValueError( + f"Detected some but not all shards of {prefix} " + "are quantized. All shards of fused layers " + "to have the same precision." + ) + else: + is_skipped = prefix in ignored_layers + + assert is_skipped is not None + return is_skipped + + +def get_pack_factor(num_bits): + assert 32 % num_bits == 0, f"Unsupported num_bits = {num_bits}" + return 32 // num_bits + + +def permute_rows( + q_w: torch.Tensor, + w_ref: torch.Tensor, + group_size: int, + test_perm: Optional[torch.Tensor] = None, +): + assert q_w.shape == w_ref.shape + + orig_device = q_w.device + k_size, _ = q_w.shape + + g_idx = torch.zeros((k_size,), dtype=torch.int32) + for i in range(k_size): + g_idx[i] = i // group_size + + # Simulate act_order by doing a random permutation on K + rand_perm = test_perm if test_perm is not None else torch.randperm(k_size) + + g_idx = g_idx[rand_perm].contiguous() + q_w = q_w[rand_perm, :].contiguous() + w_ref = w_ref[rand_perm, :].contiguous() + + return ( + w_ref.to(device=orig_device), + q_w.to(device=orig_device), + g_idx.to(device=orig_device), + rand_perm.to(device=orig_device), + ) + + +def quantize_weights( + w: torch.Tensor, + quant_type: ScalarType, + group_size: Optional[int], + zero_points: bool = False, + ref_zero_points_after_scales: bool = False, +): + assert ( + quant_type.is_integer() + ), "Floating point quantization may work but has not been tested" + assert not zero_points or group_size is not None, ( + "to have group zero points, group_size must be provided " + "(-1 group_size is channelwise)" + ) + + orig_device = w.device + orig_type = w.dtype + size_k, size_n = w.shape + + assert w.is_floating_point(), "w must be float" + + if group_size == -1: + group_size = size_k + + # Reshape to [groupsize, -1] + if group_size is not None and group_size < size_k: + w = w.reshape((-1, group_size, size_n)) + w = w.permute(1, 0, 2) + w = w.reshape((group_size, -1)) + + # Compute scale for each group + max_val = torch.max(w, 0, keepdim=True).values + min_val = torch.min(w, 0, keepdim=True).values + + max_q_val = quant_type.max() + min_q_val = quant_type.min() + + w_s = torch.Tensor([1.0]).to(w.device) # unscaled case + maybe_w_zp = None + if group_size is not None: + if zero_points: + assert not quant_type.is_signed() and quant_type.max() > 0 + w_s = (max_val - min_val).clamp(min=1e-5) / quant_type.max() + maybe_w_zp = ( + torch.round(torch.abs(min_val / w_s)).clamp(min_q_val, max_q_val).int() + ) + else: + # If the bias is such that there are no possible negative/positive + # values, set the max value to inf to avoid divide by 0 + w_s = torch.max( + abs(max_val / (max_q_val if max_q_val != 0 else torch.inf)), + abs(min_val / (min_q_val if min_q_val != 0 else torch.inf)), + ) + + # Quantize + w_q = torch.round(w / w_s).int() + (maybe_w_zp if zero_points else 0) + w_q = torch.clamp(w_q, min_q_val, max_q_val) + + # Compute ref (dequantized) + # For some kernels (namely Machete) the zero-points are applied after the + # scales are applied, for this case computing the reference in similar way + # allows us to use tighter error tolerances in our unit tests. + if ref_zero_points_after_scales and maybe_w_zp is not None: + w_ref = w_q.to(orig_type) * w_s - maybe_w_zp.to(orig_type) * w_s + else: + w_ref = (w_q - (maybe_w_zp if zero_points else 0)).to(orig_type) * w_s + + if quant_type.has_bias(): + w_q += quant_type.bias + + # Restore original shapes + if group_size is not None and group_size < size_k: + + def reshape_w(w): + w = w.reshape((group_size, -1, size_n)) + w = w.permute(1, 0, 2) + w = w.reshape((size_k, size_n)).contiguous() + return w + + w_q = reshape_w(w_q) + w_ref = reshape_w(w_ref) + w_s = w_s.reshape((-1, size_n)).contiguous() + + if maybe_w_zp is not None: + maybe_w_zp = maybe_w_zp.reshape((-1, size_n)).contiguous() + maybe_w_zp = maybe_w_zp.to(device=orig_device) + + return ( + w_ref.to(device=orig_device), + w_q.to(device=orig_device), + w_s if group_size is not None else None, + maybe_w_zp, + ) + + +def gptq_quantize_weights( + w: torch.Tensor, + quant_type: ScalarType, + group_size: int, + act_order: bool, + test_perm: Optional[torch.Tensor] = None, +): + size_k, _ = w.shape + + assert w.is_floating_point(), "w must be float" + assert ( + quant_type in SUPPORTED_GPTQ_QUANT_TYPES + ), f"Unsupported gptq type = {quant_type}" + assert group_size in SUPPORTED_GROUP_SIZES + [ + size_k + ], f"Unsupported groupsize = {group_size}" + + w_ref, w_q, w_s, _ = quantize_weights(w, quant_type, group_size) + + # Apply act_order + g_idx = torch.empty(0, dtype=torch.int, device=w.device) + rand_perm = torch.empty(0, dtype=torch.int, device=w.device) + if act_order: + assert ( + group_size < size_k + ), "For act_order, groupsize = {} must be less than size_k = {}".format( + group_size, size_k + ) + + w_ref, w_q, g_idx, rand_perm = permute_rows(w_q, w_ref, group_size, test_perm) + + return w_ref, w_q, w_s, g_idx, rand_perm + + +# QQQ employs different quant schemes for per-group and +# per-channel quantization. +def qqq_quantize_weights(w: torch.Tensor, num_bits: int, group_size: int): + orig_device = w.device + size_k, size_n = w.shape + + assert w.is_floating_point(), "w must be float" + assert ( + num_bits in MARLIN_QQQ_SUPPORTED_NUM_BITS + ), f"Unsupported num_bits = {num_bits}" + assert group_size in SUPPORTED_GROUP_SIZES + [ + size_k + ], f"Unsupported groupsize = {group_size}" + + if group_size == -1: + group_size = size_k + assert group_size <= size_k + + if group_size < size_k: + # Reshape to [groupsize, -1] + w = w.reshape((-1, group_size, size_n)) + w = w.permute(1, 0, 2) + w = w.reshape((group_size, -1)) + + max_q_val = 2**num_bits - 1 + half_q_val = (max_q_val + 1) // 2 + + # Compute scale for each group + s_group = torch.max(torch.abs(w), 0, keepdim=True)[0] + s_group *= 2 / max_q_val # 2 => symmetric + + # Quantize + q_w = torch.round(w / s_group).int() + q_w += half_q_val + q_w = torch.clamp(q_w, 0, max_q_val) + # Compute ref (dequantized) + w_ref = (q_w - half_q_val).half() * s_group + + # Restore original shapes + def reshape_w(w): + w = w.reshape((group_size, -1, size_n)) + w = w.permute(1, 0, 2) + w = w.reshape((size_k, size_n)).contiguous() + return w + + q_w = reshape_w(q_w) + w_ref = reshape_w(w_ref) + + # Compute int8 quantization scale for each channel + s_channel = torch.max(torch.abs(w_ref), 0, keepdim=True)[0] + s_channel /= 127.0 + t_int8 = (w_ref / s_channel).round().clamp(-128, 127).to(torch.int8) + w_ref = t_int8.half() * s_channel + s_channel = s_channel.reshape(1, -1).to(dtype=torch.float) + + # Fuse scales + s_group = (s_group.reshape(-1, size_n).contiguous() / s_channel).to( + dtype=torch.half + ) + else: + max_q_val = 2 ** (num_bits - 1) - 1 + + # Compute scale for each channel + s_channel = torch.max(torch.abs(w), 0, keepdim=True)[0] + s_channel /= max_q_val + + # Quantize + q_w = torch.round(w / s_channel).int() + q_w = torch.clamp(q_w, -max_q_val, max_q_val) + # Compute ref (dequantized) + w_ref = q_w.half() * s_channel + + s_group = torch.tensor([], dtype=torch.half) + # div 2 ** (8 - self.bits)) to offset right shift in unpacking + s_channel /= 2 ** (8 - num_bits) + s_channel = s_channel.reshape(-1, size_n).contiguous().to(torch.float) + + return ( + w_ref.to(device=orig_device), + q_w.to(device=orig_device), + s_group.to(device=orig_device), + s_channel.to(device=orig_device), + ) + + +def sort_weights(q_w: torch.Tensor, g_idx: torch.Tensor): + orig_device = q_w.device + + sort_indices = torch.argsort(g_idx).to(dtype=torch.int32) # Sort based on g_idx + + g_idx = g_idx[sort_indices].contiguous() + q_w = q_w[sort_indices, :].contiguous() + + return ( + q_w.to(device=orig_device), + g_idx.to(device=orig_device), + sort_indices.to(device=orig_device), + ) + + +def pack_rows( + q_w: torch.Tensor, + num_bits: int, + size_k: int, + size_n: int, +): + assert q_w.shape == (size_k, size_n) + + pack_factor = get_pack_factor(num_bits) + assert size_k % pack_factor == 0 + + orig_device = q_w.device + + q_w = q_w.cpu().numpy().astype(numpy.uint32) + + q_res = numpy.zeros((size_k // pack_factor, size_n), dtype=numpy.uint32) + + for i in range(pack_factor): + q_res |= q_w[i::pack_factor, :] << num_bits * i + + q_res = torch.from_numpy(q_res.astype(numpy.int32)).to(orig_device) + return q_res + + +def pack_cols( + q_w: torch.Tensor, + num_bits: int, + size_k: int, + size_n: int, +): + assert q_w.shape == (size_k, size_n) + + pack_factor = get_pack_factor(num_bits) + assert size_n % pack_factor == 0 + + orig_device = q_w.device + + q_w = q_w.cpu().numpy().astype(numpy.uint32) + + q_res = numpy.zeros((size_k, size_n // pack_factor), dtype=numpy.uint32) + + for i in range(pack_factor): + q_res |= q_w[:, i::pack_factor] << num_bits * i + + q_res = torch.from_numpy(q_res.astype(numpy.int32)).to(orig_device) + q_res = q_res.contiguous() + + return q_res + + +def unpack_cols( + packed_q_w: torch.Tensor, + num_bits: int, + size_k: int, + size_n: int, +): + pack_factor = get_pack_factor(num_bits) + assert size_n % pack_factor == 0 + assert packed_q_w.shape == ( + size_k, + size_n // pack_factor, + ), "packed_q_w.shape = {} size_k = {}, size_n = {} pack_Factor = {}".format( + packed_q_w.shape, size_k, size_n, pack_factor + ) + + orig_device = packed_q_w.device + + packed_q_w_cpu = packed_q_w.cpu().numpy().astype(numpy.uint32) + q_res = numpy.zeros((size_k, size_n), dtype=numpy.uint32) + + mask = (1 << num_bits) - 1 + for i in range(pack_factor): + vals = packed_q_w_cpu & mask + packed_q_w_cpu >>= num_bits + q_res[:, i::pack_factor] = vals + + q_res = torch.from_numpy(q_res.astype(numpy.int32)).to(orig_device) + q_res = q_res.contiguous() + + return q_res + + +def gptq_pack( + q_w: torch.Tensor, + num_bits: int, + size_k: int, + size_n: int, +): + return pack_rows(q_w, num_bits, size_k, size_n) + + +def awq_pack( + q_w: torch.Tensor, + num_bits: int, + size_k: int, + size_n: int, +): + assert q_w.shape == (size_k, size_n) + + # Interleave column dim (for the dequantize code) and pack it to int32 + if num_bits == 4: + interleave = numpy.array([0, 2, 4, 6, 1, 3, 5, 7]) + elif num_bits == 8: + interleave = numpy.array([0, 2, 1, 3]) + else: + raise Exception("num_bits must be 4 or 8, got {}".format(num_bits)) + + q_w = q_w.reshape((-1, len(interleave)))[:, interleave].ravel() + q_w = q_w.reshape((-1, size_n)).contiguous() + + return pack_cols(q_w, num_bits, size_k, size_n) diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/__init__.py b/build/torch27-cxx11-cu128-aarch64-linux/moe/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..0d9d8227711def6d51c24b5fd779a3d13681d88c --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/__init__.py @@ -0,0 +1,80 @@ +import torch + +from . import layers +from ._ops import ops +from .fp8_utils import per_token_group_quant_fp8, w8a8_block_fp8_matmul +from .fused_marlin_moe import fused_marlin_moe +from .fused_moe import fused_experts, fused_moe, fused_topk, grouped_topk +from .scalar_type import ScalarType, scalar_types + + +def gptq_marlin_moe_repack( + b_q_weight: torch.Tensor, + perm: torch.Tensor, + size_k: int, + size_n: int, + num_bits: int, +) -> torch.Tensor: + num_experts = b_q_weight.shape[0] + assert size_k % 16 == 0 + output = torch.empty( + (num_experts, size_k // 16, size_n * (num_bits // 2)), + device=b_q_weight.device, + dtype=b_q_weight.dtype, + ) + for e in range(num_experts): + output[e] = ops.gptq_marlin_repack( + b_q_weight[e], perm[e], size_k, size_n, num_bits + ) + return output + + +def awq_marlin_moe_repack( + b_q_weight: torch.Tensor, + perm: torch.Tensor, + size_k: int, + size_n: int, + num_bits: int, +) -> torch.Tensor: + num_experts = b_q_weight.shape[0] + assert size_k % 16 == 0 + output = torch.empty( + (num_experts, size_k // 16, size_n * (num_bits // 2)), + device=b_q_weight.device, + dtype=b_q_weight.dtype, + ) + for e in range(num_experts): + output[e] = ops.awq_marlin_repack(b_q_weight[e], size_k, size_n, num_bits) + return output + + +def moe_sum(input: torch.Tensor, output: torch.Tensor): + ops.moe_sum(input, output) + + +def topk_softmax( + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + token_expert_indicies: torch.Tensor, + gating_output: float, +) -> None: + ops.topk_softmax(topk_weights, topk_ids, token_expert_indicies, gating_output) + + +__all__ = [ + "ScalarType", + "awq_marlin_moe_repack", + "fused_experts", + "fused_marlin_moe", + "fused_moe", + "fused_topk", + "gptq_marlin_moe_repack", + "grouped_topk", + "layers", + "moe_align_block_size", + "moe_sum", + "per_token_group_quant_fp8", + "scalar_types", + "topk_softmax", + "w8a8_block_fp8_matmul", +] diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/_moe_69e0359.abi3.so b/build/torch27-cxx11-cu128-aarch64-linux/moe/_moe_69e0359.abi3.so new file mode 100755 index 0000000000000000000000000000000000000000..a8071b8e60b3d805d536c329ac5007f72b7b8cfc --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/_moe_69e0359.abi3.so @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8c5f27c5de9eef7da26b8e1eeccdd30f35922ff4d04da2ef16786e76682a85be +size 119621992 diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/_ops.py b/build/torch27-cxx11-cu128-aarch64-linux/moe/_ops.py new file mode 100644 index 0000000000000000000000000000000000000000..b33dfcc6afacf69f6989c184c5c3efabc03f18da --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/_ops.py @@ -0,0 +1,9 @@ +import torch +from . import _moe_69e0359 +ops = torch.ops._moe_69e0359 + +def add_op_namespace_prefix(op_name: str): + """ + Prefix op by namespace. + """ + return f"_moe_69e0359::{op_name}" \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..56c1a4e3af0b4a93fff71028d8e04bf73f0abb29 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..d3677bebb82a7f3f19344ef6471626493cf2c5bb --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=1,N=1792,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=1,N=1792,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..265768fb900ccfe9612b4a0d25973e6618f22a79 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=1,N=1792,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "5120": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "9216": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "13312": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "17408": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "25600": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "33792": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "41984": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "50176": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "58368": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=1,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=1,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..d3be23dfc903ba61d3d4d79c0230952b24d2ead0 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=1,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "5120": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "9216": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "13312": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "17408": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "25600": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "33792": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "41984": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "50176": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "58368": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=1,N=3072,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=1,N=3072,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..589f5d39f31418d5121e7cbb2e6f2894b0a7ed32 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=1,N=3072,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "5120": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "9216": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "13312": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "17408": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "25600": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "33792": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "41984": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "50176": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "58368": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=1,N=3072,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=1,N=3072,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..2c78bfaba7890772bf266721f5577202ea443882 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=1,N=3072,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "5120": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "9216": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "13312": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "17408": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "25600": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "33792": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "41984": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "50176": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "58368": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=1,N=3072,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=1,N=3072,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..4da841e74a79f9589fecac1fa557ea132d34805f --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=1,N=3072,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "5120": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "9216": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "13312": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "17408": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "25600": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "33792": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "41984": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "50176": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "58368": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=1,N=3584,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=1,N=3584,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..200356713c0d0a76e199671c7ec8f10d0e5ee0ac --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=1,N=3584,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "5120": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "9216": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "13312": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "17408": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "25600": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "33792": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "41984": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "50176": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "58368": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=1,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=1,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..e076615ee541a5043556f630ecf0946c4e2c1408 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=1,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "5120": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "9216": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "13312": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "17408": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "25600": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "33792": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "41984": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "50176": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "58368": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=1,N=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=1,N=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..ee896554b921040d7810bb6e9368cc200777951d --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=1,N=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "5120": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "9216": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "13312": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "17408": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "25600": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "33792": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "41984": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "50176": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "58368": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=1,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=1,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..05aed8b1c81492151d128ef251afc510d8cc8ed5 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=1,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "5120": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "9216": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "13312": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "17408": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "25600": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "33792": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "41984": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "50176": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "58368": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=1344,device_name=NVIDIA_A100-SXM4-40GB.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=1344,device_name=NVIDIA_A100-SXM4-40GB.json new file mode 100644 index 0000000000000000000000000000000000000000..9262a74a4a0e1e3789f260a3ef7f6cb9551f3f2b --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=1344,device_name=NVIDIA_A100-SXM4-40GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=1344,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=1344,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..d251f9b5accaec977fc87a0999cd56ee387fc650 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=1344,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=1344,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=1344,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..0ecf814a28a9441e89f892eb3d63dcf8dcb0dd97 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=1344,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=14336,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=14336,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..51ad5b299eb22465fa80530d12bdd5d7a03ce398 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=14336,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=14336,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=14336,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..ee5119182556cf49434c10e56cf04e3baeb26408 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=14336,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=1792,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=1792,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..68793c77b33c4f4b97d0a4b780fcbe8043c799de --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=1792,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "5120": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "9216": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "13312": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "17408": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "25600": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "33792": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "41984": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "50176": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "58368": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..612910720ed9439e56c4af4c03f30fee224fac80 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "5120": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "9216": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "13312": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "17408": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "25600": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "33792": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "41984": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "50176": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "58368": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=2688,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=2688,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..039a10ed127b77836a7f41c03513292613852b30 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=2688,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=2688,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=2688,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..3793fcafee60bc7e8f5f12d601cb3192abfa9ca8 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=2688,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=3072,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=3072,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..51d03d8607122d7b9bc20ba48d8432d62367fa00 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=3072,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=3072,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=3072,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..26f9abd6b789e9dd0f83ec7721fd1bae8aa76bec --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=3072,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "1536": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=3200,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=3200,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..cd0cdbea0c3372674cb610870dd0b30325864549 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=3200,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,130 @@ +{ + "3328": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "768": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "1792": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2560": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "2816": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "3584": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "3840": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1280": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "2304": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=3584,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=3584,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..64be6e6591422aa0f441c3747b6c49850929652e --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=3584,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..0a6a6a73fa45e270f01ba7ebdc6d9d55bf9daad3 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,218 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "5120": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "9216": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "13312": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "17408": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "25600": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "33792": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "41984": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "50176": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "58368": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=6400,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=6400,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..ba9041d008507e31ae4179ef2bc863a49c606582 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=6400,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,130 @@ +{ + "3840": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "1792": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "3584": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "2816": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1280": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "768": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "3328": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "2560": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "2304": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..7a7508aab04599cb06641c835d8b0a14f54d0716 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..dbf9a2dd6f048d8adee290961e2aea72035f7615 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json new file mode 100644 index 0000000000000000000000000000000000000000..bbb2386046b1135a2cc7ab7cb26c1d0b039bcf3a --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "1536": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=800,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=800,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..57055453aa24c831dad9ac8e37fdab707c63ef91 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=16,N=800,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,130 @@ +{ + "2048": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "1792": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "3328": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "2560": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 4 + }, + "768": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "2816": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "2304": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2 + }, + "1280": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "3840": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "3584": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=160,N=192,device_name=NVIDIA_A800-SXM4-80GB.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=160,N=192,device_name=NVIDIA_A800-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..0611620eb33626ac891454637723eabec9aa1183 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=160,N=192,device_name=NVIDIA_A800-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=1024,device_name=AMD_Instinct_MI325X,block_shape=[128,128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=1024,device_name=AMD_Instinct_MI325X,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..43c249d2530e1fafb46f6594b73e11f33fa8ffe7 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=1024,device_name=AMD_Instinct_MI325X,block_shape=[128,128].json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=1024,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=1024,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..43c249d2530e1fafb46f6594b73e11f33fa8ffe7 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=1024,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..4dd00d110e486c0a7670504ace778fd5995de727 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..48f9697af2639bd4afd45c768c7a8e9442924e2d --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..a8c05712ba5872db5414e571fef659572e59eb80 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..f1244c61efb019183e892f848700072c3be03ef9 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..2e692a1583a4ad95f020634d55f7f4e327be5da0 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128,128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..a2ee05da1d7c632aa6b1c3ccc8af2546f987ee93 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=128,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=256,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=256,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..63e118746fd86305fd03064637ad915b72c8abcd --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=256,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=256,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=256,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..e6769604ee6bcc75c41a5bea296cb55da3c755e8 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=256,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=256,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=256,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..e6769604ee6bcc75c41a5bea296cb55da3c755e8 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=256,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..fc573cd6e8561f56830c8f8ca3033f2d2f7bc1b3 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128,128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..c6d7e96c7f0aea61a2bd57723ecb91b1c1230895 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128,128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..6fcf408755f5d3e28ff636472f89e4fa52c8da22 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_L20,dtype=fp8_w8a8,block_shape=[128,128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_L20,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..c6eabea66a39aa86f5becc04d32ee01dbe627dce --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=256,device_name=NVIDIA_L20,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=512,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=512,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json new file mode 100644 index 0000000000000000000000000000000000000000..e6769604ee6bcc75c41a5bea296cb55da3c755e8 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=512,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=64,device_name=NVIDIA_A800-SXM4-80GB.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=64,device_name=NVIDIA_A800-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..21f60229ff875ca6dc0cd0a3cb4fa51a845bf881 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=256,N=64,device_name=NVIDIA_A800-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=60,N=1408,device_name=AMD_Instinct_MI300X.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=60,N=1408,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..d09508b31729ddbafc8b5e1dfa88343c61e83ca9 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=60,N=1408,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "8": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=60,N=176,device_name=AMD_Instinct_MI300X.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=60,N=176,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..746463af4d56315af3c31348c2253e43c926553e --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=60,N=176,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "3072": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=60,N=352,device_name=AMD_Instinct_MI300X.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=60,N=352,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..bbdb9ad096451d6c5b9d95fae4ed02ce37715002 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=60,N=352,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "2048": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "3072": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=60,N=704,device_name=AMD_Instinct_MI300X.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=60,N=704,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..43584b1eb6b65f02fc919490b0ff4b9569de58c4 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=60,N=704,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..8cc6c643f236d2f7f9ad29354d9e469d00b20d3f --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_A800-SXM4-80GB.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_A800-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..39a9912fa4bddc48de38d5a8b4fadd71326485d2 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_A800-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..05b54639d234e20096744915c932bb2d415641a5 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..d4c9ddd12972ac0b5fd2be11a9cd1075906e3978 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..c17a4ec346915af69f10f65410450cbd66614810 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H200.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H200.json new file mode 100644 index 0000000000000000000000000000000000000000..170ae7f3fff1d09961d368853dcfc2f24c1b878c --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=1280,device_name=NVIDIA_H200.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=2560,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=2560,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..1d9d352edebc3f228353873677d2369e6a721158 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=2560,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=2560,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=2560,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..9ad5b316750051e4a5dc39b195f6a29eeaa882f0 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=2560,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=2560,device_name=NVIDIA_H200.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=2560,device_name=NVIDIA_H200.json new file mode 100644 index 0000000000000000000000000000000000000000..2883dfd11e7f3a7db429ae0aaeb98d554f271cdf --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=2560,device_name=NVIDIA_H200.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..8abfd84a776b7be2c0920ca8d3a69c06735a08ee --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..2fc18a5e43d298ddabc57dc9f357296890fb7e29 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..be8d4a7fd23d994ac838d45bf249e39764552ccf --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H200.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H200.json new file mode 100644 index 0000000000000000000000000000000000000000..71fdd88643c6ff515808fdc2d67e9fda3ac62ae9 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=320,device_name=NVIDIA_H200.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..b2799ed3a866e25b78d60d92910c000ebb21ff71 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_A800-SXM4-80GB.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_A800-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..c02de2f628b716456373978f600f9fc8c4b98d24 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_A800-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_GeForce_RTX_4090,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_GeForce_RTX_4090,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..3e0bc75ff87c43ad3a0b3f132e5367a8ef39619a --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_GeForce_RTX_4090,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..9f7ed6726f44e60dea5af588e441b269b6c5cc3d --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..b8d3be2313fa14025d8aeb2fd11e0d1ee997ffa6 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..21b72557e365d01011f48a046915765e67dbd13f --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H200.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H200.json new file mode 100644 index 0000000000000000000000000000000000000000..eaf32f6d76c0a6aa50a5fbc2f055a5e7c9e10b66 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=64,N=640,device_name=NVIDIA_H200.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..b6f1d01f886524c58932ac28c31dcf6f10788e14 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..4bf775347eccb3e76e20ea0018d71343f9848313 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..f245285bd821717f2280fa89b5f0aa6bbce2262d --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI325X.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI325X.json new file mode 100644 index 0000000000000000000000000000000000000000..3918c93b160a39e55153cbf6ff634d59e724d021 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI325X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=14336,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=14336,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..3f3ccdafa88f3452a695efad4cb9622d6ae79e6a --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=14336,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,138 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=14336,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=14336,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..841044a4fc6e2da81f18b2765fa30b9de7092b3d --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=14336,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=14336,device_name=NVIDIA_H200.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=14336,device_name=NVIDIA_H200.json new file mode 100644 index 0000000000000000000000000000000000000000..59be497fc428755f010e490277f4b945d6ed0195 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=14336,device_name=NVIDIA_H200.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..0e5fd1eec77d7e0aca0f1e3376b2c45217ce234a --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI300X.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..d6ad63509f1573d25fbdc370b93bc39a0d00b51a --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..16e0a91baf31d9a819f27eeea719e564f399379c --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI325X.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI325X.json new file mode 100644 index 0000000000000000000000000000000000000000..d766fc062ddcce10d38bdc68ab1f700d193b8ba7 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI325X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..8323f512db0158570427a7eb0a69bd889bc2d9ed --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI300X.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..1b46cb5716514f290c6b4bc3a18dba490c799272 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..6d5b1ae5b15fc333d3cdaa5dbcacab33267339cc --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI325X.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI325X.json new file mode 100644 index 0000000000000000000000000000000000000000..ffc1b23ea90ddefd24b4b98a222e67e84442cf9c --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI325X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_A100-SXM4-40GB.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_A100-SXM4-40GB.json new file mode 100644 index 0000000000000000000000000000000000000000..f4c0f8417b384870050a95e0cf57edbdf6352b23 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_A100-SXM4-40GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..5c8185cfdeec167ec4b88de51b4b395e28769cc5 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..97c9f4445b166657ad29f1db9fc8281f9c463ec4 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..e4110a5d2e70f4d2698d4c28b1668f8e4a48688a --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_H200.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_H200.json new file mode 100644 index 0000000000000000000000000000000000000000..0883ef40582ea4f1b08c629edcd6dd9beaaf2dad --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=1792,device_name=NVIDIA_H200.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..81bb765d30031ca3031636d7a442dc2c2292a6a9 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI300X.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..811c77ab4109319c9f079e414bd4f7d2cf4ac9f1 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..2758e48fc4065cd05ceb6fe22574acf2b3aae633 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI325X.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI325X.json new file mode 100644 index 0000000000000000000000000000000000000000..fc31215cbae8afeee27d923e176a25daf1f2b43d --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI325X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..0bb423b28f5ab3825929a4870b96393262a9dd9f --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..55571873395464a3b58f549523905f439a8f1716 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..26bcbf26970c7a77c99e2c8eacd83eefa86967bf --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..1a0aa331933291f0c32b2cf26672771b5c7b4196 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H200.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H200.json new file mode 100644 index 0000000000000000000000000000000000000000..9952be6ba4abe434e2924e9505e834de4ce0528e --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=2048,device_name=NVIDIA_H200.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..379ca107a9469a5e5777e4e648f3f25ffc382a0a --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..5a3f415d54142383fa7677d8192f7a4c4df66b2e --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..6cb80f48329f5718e892fd6f78000df4003832f2 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI325X.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI325X.json new file mode 100644 index 0000000000000000000000000000000000000000..de9d0aba75a7ed45fdfd92c123736f1e981a06eb --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI325X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_A100-SXM4-40GB.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_A100-SXM4-40GB.json new file mode 100644 index 0000000000000000000000000000000000000000..b41f9d443e50678334f906b44fce6d018d69500e --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_A100-SXM4-40GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..edf2a38d12ad3f420f232d2cd61ab149ad138725 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_GeForce_RTX_4090,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_GeForce_RTX_4090,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..32bbadbb9eae8d36e30895d863a8d10ff65167b2 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_GeForce_RTX_4090,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..673bae2ba8ef80ed4d4930739ca7daf0e8f28ee1 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..b2100cebb7f589747430be9ca8c8db368c152d78 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..e6f753cdba35b55e8d021597dc4911bab27caa8d --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H200.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H200.json new file mode 100644 index 0000000000000000000000000000000000000000..53f3394693f060daebff9960a95ed3753e8522c9 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_H200.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_L40S.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_L40S.json new file mode 100644 index 0000000000000000000000000000000000000000..d720deb4bdd73d194b1023c99e190b8fcfecdaef --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=3584,device_name=NVIDIA_L40S.json @@ -0,0 +1,173 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 2 + }, + "2": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 2, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 7 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 128, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 1 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 1 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 2, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 2, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 2, + "num_warps": 4, + "num_ctas": 1, + "num_stages": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 2 + }, + "192": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 8, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_ctas": 1, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 128, + "num_warps": 2, + "num_ctas": 1, + "num_stages": 8 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_ctas": 1, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_ctas": 1, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 2 + }, + "6144": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_ctas": 1, + "num_stages": 2 + }, + "8192": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 16, + "num_ctas": 1, + "num_stages": 2 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..48bb5f2ccb8e3c623593538461d8d8ba71f93a69 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI300X.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..a64d06c6d1724f736c5081a8b947fa2e6127701d --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..2c49f359c22ac8d3b8fd3b00035b70472c34ed51 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI325X.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI325X.json new file mode 100644 index 0000000000000000000000000000000000000000..c7db6c0cbd3faf615074fc8055ff744e528dd204 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI325X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..dbc624731f5cb9afcdc9213183d00d1e5edd4a00 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..cc614e635ea57327c610ce79e99ae5339614f22e --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..32c0c9da471cbe479044095e0ed14a0f54b73620 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..4dd475c02a19b380efbe8f59945a0f2a41acda11 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H200.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H200.json new file mode 100644 index 0000000000000000000000000000000000000000..2ed15f30fe603c080a2e8a32b664400a50d2153e --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H200.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..bd2c6fbc1b941cebfe8e03f62e3044f5878db0ed --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..8d7b780271854c0fbae29e92cfd3cce508ca029e --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..7a07bbf4141952c4dc6199676b1ef97d1cfabece --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI325X.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI325X.json new file mode 100644 index 0000000000000000000000000000000000000000..3a3268cc17a84d88f0a67e20e266491ad45d1777 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI325X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 32, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json new file mode 100644 index 0000000000000000000000000000000000000000..f578c8d0160ac3ef85b53c8539d3675455a97173 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..918f6839620cbab1f30b0f9383a9129c2cf2cf3d --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H100_80GB_HBM3.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 0000000000000000000000000000000000000000..e341a67917d5177bacb3f6767e7b6d92539826ad --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..eb817268d41208a661d2ab9526c5108d36d44085 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H200.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H200.json new file mode 100644 index 0000000000000000000000000000000000000000..0c7062aea6c4e920cdfdc9e13fc2c0fa38aceb0e --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=7168,device_name=NVIDIA_H200.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..cd4fb8f11b935e401103aab5117a1d1be1799c6e --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI300X.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI300X.json new file mode 100644 index 0000000000000000000000000000000000000000..cf66868e9d57a541fe10fa9fff682f7a7c6dac97 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI300X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 1, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..c27ca0a36594fd12a9995fffda60a9d5003425c0 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json @@ -0,0 +1,164 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI325X.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI325X.json new file mode 100644 index 0000000000000000000000000000000000000000..da477b1fb15eb2064cbff78b5e47ae9c54dae7a5 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI325X.json @@ -0,0 +1,200 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 16, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 2, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 1 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 4, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 4, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2, + "waves_per_eu": 0, + "matrix_instr_nonkdim": 16, + "kpack": 2 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=8192,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=8192,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..34b916e574f88c65db1dac5889d74a990dc25e9b --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=8192,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=8192,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=8192,device_name=NVIDIA_H200,dtype=fp8_w8a8.json new file mode 100644 index 0000000000000000000000000000000000000000..96cbc111c7fffe8b0a67c22711f9a5c8f28d8b61 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/E=8,N=8192,device_name=NVIDIA_H200,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} \ No newline at end of file diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=1536,K=1536,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=1536,K=1536,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..6496a38fba8ae09b3025a75f357815b9d6a5e3f4 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=1536,K=1536,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=1536,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=1536,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..3618053b65831b95c4bb0f20ef3b9aa816b2d637 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=1536,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=1536,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=1536,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..46a982f5ee9a4bd67ce244b101c576efeeb53b78 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=1536,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=2048,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=2048,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..035ec027fa56622196b24a03a5042ce010deaebf --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=2048,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=2048,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=2048,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..8b49f2781cb54d19a2789767ebb7e8c3fb55b981 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=2048,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=2304,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=2304,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..851bc9f9f0b50b41451b929eaa518869b6a05412 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=2304,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=2304,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=2304,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..d1227c2157990216d2ca51c69ad0944017f53b6a --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=2304,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=24576,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=24576,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..77ba0d7477bdbcb036a43263e7aaa6b6913f8f4e --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=24576,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=24576,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=24576,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..1c61451fb34e52deec827f8f63c80fb15830c202 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=24576,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=24576,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=24576,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..63e661c80de6a7b1422f7a994a2ee7a4b724911c --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=24576,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=256,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=256,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..cf354037903c0d1fcd077c4647aabce026a723fb --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=256,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=3072,K=1536,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=3072,K=1536,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..0a5d7bfdba4852da9ed08d1bc27cd7d521d09965 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=3072,K=1536,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=3072,K=1536,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=3072,K=1536,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..eccb86a76df0d7302b760ab6d83a8ceb9fa9d0d9 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=3072,K=1536,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=3072,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=3072,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..cb91a279d423d0ca25197e0edd5e8c2f4da58720 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=3072,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=3072,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=3072,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..88af48431d8b8791af8df03429704606b670f1f7 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=3072,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=3072,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=3072,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..dd069726d7ed4dcbb449af243f4f4af21815f854 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=3072,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=32768,K=512,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=32768,K=512,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..7febe3d272b4bb76500f7c6b523396129fd53680 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=32768,K=512,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=32768,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=32768,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..56b939e52fac3ed53a4e0ba640c40010cb3af30a --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=32768,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=32768,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=32768,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..63d9a0bf5d79ddaaad547d44338ad4b959ad72b1 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=32768,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=36864,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=36864,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..7fa398c15a2a535401709b0f25e20f6e4b23e58e --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=36864,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=36864,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=36864,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..f15d8f64c7090bd71d0091a524c65d7818fec38e --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=36864,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=4096,K=512,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=4096,K=512,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..9d7658bfc41b2c8fd4daf3fbdf62d15936d3d546 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=4096,K=512,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=4096,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=4096,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..cd3e07804fdec10c2cfb291c1ede3ba67b753f9c --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=4096,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=4096,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=4096,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..9d5a329d7466a37c0ca68a65a089fbb99f9327a9 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=4096,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=4608,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=4608,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..03dba5ad15ba5f7f49100a5c78e8685e64334b2a --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=4608,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=4608,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=4608,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..96e1594a3eabbaedc792b84b07f05ae8752b7251 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=4608,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=4608,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=4608,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..5ffd367df833d773355590220598a3c7eceba4e0 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=4608,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=512,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=512,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..9a5ff48b8942957dde9b862aed848390dd267948 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=512,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=512,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=512,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..eabc423949a24c2a1fb2368a73e5249caf8d07df --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=512,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=576,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=576,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..386928de139ce718f28222b9c1a6555df3958491 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=576,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 5 + }, + "512": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 256, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=576,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=576,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..51e237b91b8e775a36bcf783c078c2c1cecbcbd2 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=576,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=576,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=576,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..6280219c9ee7d26f7e2fd3625dc92d847ddc7982 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=576,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=1024,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=1024,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..40c01c0b92b4b26fe480879dda33f18c5eb59a6d --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=1024,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=1024,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=1024,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..c6fd3659799bc31e17f3577e7f0e8d7268faf1fb --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=1024,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=1152,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=1152,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..160f12ed3f95a6967439ff53bc3e3a2cdc97c700 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=1152,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=1152,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=1152,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..e5c4a1d2c94e5c7864f462e083ea5f530b8efe3f --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=1152,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=128,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=128,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..2bf5eb27e38208871d50348b170c8c74b80fc519 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=128,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 2 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "96": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=16384,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=16384,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..9c908e80406587da4d246ce4e3a8a98a14c875b1 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=16384,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=16384,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=16384,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..0a1e14cffbb2a894a701352193947d272427db0d --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=16384,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=16384,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=16384,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..15b1c93f60fc5068ba11b82b6d5924dd2024a824 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=16384,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=18432,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=18432,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..f78e7060e6840ff721d306db556636b0bbc8d9b3 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=18432,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=18432,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=18432,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..8ff12e64c172f5a5d0fbdf900728fe60b33877e2 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=18432,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=18432,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=18432,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..4532f93681e2be175b1bf94f81bfde711821cd60 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=18432,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=2048,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=2048,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..1d3ce5c94c2d9a4a1637204efb3b14f7a5579bdb --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=2048,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=2048,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=2048,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..ca7f32b9552b479dc05495792b7e426db5eb1b56 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=2048,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=2048,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=2048,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..5acea242cc0ad094cba8ee5f568ff88afb1b41ae --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=2048,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=2304,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=2304,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..3ab5796ee15b6ec8d4ab1f4ab5a594fecb30e4b4 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=2304,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "24": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=2304,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=2304,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..58cdd93e90b8c29bc7a211861711565dbeeb529a --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=2304,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=2304,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=2304,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..b72e0371d1421a1decc9d57860f83eea8f790942 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=2304,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=256,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=256,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..3cb7eaa07c745fd3aa2b3242780a7061bedac1de --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=256,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 2 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 5 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2 + }, + "1536": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json new file mode 100644 index 0000000000000000000000000000000000000000..293adce387e066fce75b6e606d4b8b6a5aa10bdb --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/configs/N=7168,K=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "256": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 4 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + } +} diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/fp8.py b/build/torch27-cxx11-cu128-aarch64-linux/moe/fp8.py new file mode 100644 index 0000000000000000000000000000000000000000..23bd7d6703104b0020671cc6ba6f78a6df37e4bf --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/fp8.py @@ -0,0 +1,63 @@ +from typing import Optional, Tuple, Union + +import torch + +from ._ops import ops + + +def is_hip() -> bool: + return torch.version.hip is not None + + +def scaled_fp8_quant( + input: torch.Tensor, + scale: Optional[torch.Tensor] = None, + num_token_padding: Optional[int] = None, + scale_ub: Optional[torch.Tensor] = None, + use_per_token_if_dynamic: bool = False, +) -> Tuple[torch.Tensor, torch.Tensor]: + """ + Quantize input tensor to FP8 and return quantized tensor and scale. + + This function supports both static and dynamic quantization: If you + provide the scale, it will use static scaling and if you omit it, + the scale will be determined dynamically. The function also allows + optional padding of the output tensors for downstream kernels that + will benefit from padding. + + Args: + input: The input tensor to be quantized to FP8 + scale: Optional scaling factor for the FP8 quantization + scale_ub: Optional upper bound for scaling factor in dynamic + per token case + num_token_padding: If specified, pad the first dimension + of the output to at least this value. + use_per_token_if_dynamic: Whether to do per_tensor or per_token + in the dynamic quantization case. + + Returns: + Tuple[torch.Tensor, torch.Tensor]: The output tensor in FP8 and + scaling factor. + """ + # This code assumes batch_dim and num_tokens are flattened + assert input.ndim == 2 + shape: Union[Tuple[int, int], torch.Size] = input.shape + # For rocm, the output fp8 dtype is torch.float_e3m3fnuz + out_dtype: torch.dtype = torch.float8_e4m3fnuz if is_hip() else torch.float8_e4m3fn + if num_token_padding: + shape = (max(num_token_padding, input.shape[0]), shape[1]) + output = torch.empty(shape, device=input.device, dtype=out_dtype) + + if scale is None: + if use_per_token_if_dynamic: + scale = torch.empty((shape[0], 1), device=input.device, dtype=torch.float32) + ops.dynamic_per_token_scaled_fp8_quant(output, input, scale, scale_ub) + else: + scale = torch.zeros(1, device=input.device, dtype=torch.float32) + ops.dynamic_scaled_fp8_quant(output, input, scale) + else: + # num_token_padding not implemented for this case + assert scale.numel() == 1 or num_token_padding is None + ops.static_scaled_fp8_quant(output, input, scale) + + return output, scale diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/fp8_utils.py b/build/torch27-cxx11-cu128-aarch64-linux/moe/fp8_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..acb4f3e3bb1a34f209fdac9ecca8c123aaf67f12 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/fp8_utils.py @@ -0,0 +1,391 @@ +import functools +import json +import logging +import os +from typing import Any, Dict, List, Optional, Tuple + +import torch +import triton +import triton.language as tl + +from .platforms import current_platform + +logger = logging.getLogger(__name__) + + +@triton.jit +def _per_token_group_quant_fp8( + # Pointers to inputs and output + y_ptr, + y_q_ptr, + y_s_ptr, + group_size, + # Avoid to divide zero + eps, + # Information for float8 + fp8_min, + fp8_max, + # Meta-parameters + BLOCK: tl.constexpr, +): + """A Triton-accelerated function to perform per-token-group + quantization on a tensor. + This function converts the tensor values into float8 values. + """ + # Map the program id to the row of X and Y it should compute. + g_id = tl.program_id(0) + y_ptr += g_id * group_size + y_q_ptr += g_id * group_size + y_s_ptr += g_id + + cols = tl.arange(0, BLOCK) # N <= BLOCK + mask = cols < group_size + + y = tl.load(y_ptr + cols, mask=mask, other=0.0).to(tl.float32) + # Quant + _absmax = tl.maximum(tl.max(tl.abs(y)), eps) + y_s = _absmax / fp8_max + y_q = tl.clamp(y / y_s, fp8_min, fp8_max).to(y_q_ptr.dtype.element_ty) + + tl.store(y_q_ptr + cols, y_q, mask=mask) + tl.store(y_s_ptr, y_s) + + +@triton.jit +def _per_token_group_quant_fp8_colmajor( + # Pointers to inputs and output + y_ptr, + y_q_ptr, + y_s_ptr, + group_size, + # Num columns of y + y_num_columns, + # Stride from one column to the next of y_s + y_s_col_stride, + # Avoid to divide zero + eps, + # Information for float8 + fp8_min, + fp8_max, + # Meta-parameters + BLOCK: tl.constexpr, +): + """A Triton-accelerated function to perform per-token-group + quantization on a tensor. + This function converts the tensor values into float8 values. + """ + # Map the program id to the row of X and Y it should compute. + g_id = tl.program_id(0) + y_ptr += g_id * group_size + y_q_ptr += g_id * group_size + + # Convert g_id the flattened block coordinate to 2D so we can index + # into the output y_scales matrix + blocks_per_row = y_num_columns // group_size + scale_col = g_id % blocks_per_row + scale_row = g_id // blocks_per_row + y_s_ptr += scale_col * y_s_col_stride + scale_row + + cols = tl.arange(0, BLOCK) # group_size <= BLOCK + mask = cols < group_size + + y = tl.load(y_ptr + cols, mask=mask, other=0.0).to(tl.float32) + # Quant + _absmax = tl.maximum(tl.max(tl.abs(y)), eps) + y_s = _absmax / fp8_max + y_q = tl.clamp(y / y_s, fp8_min, fp8_max).to(y_q_ptr.dtype.element_ty) + + tl.store(y_q_ptr + cols, y_q, mask=mask) + tl.store(y_s_ptr, y_s) + + +def per_token_group_quant_fp8( + x: torch.Tensor, + group_size: int, + eps: float = 1e-10, + dtype: Optional[torch.dtype] = None, + column_major_scales: bool = False, +) -> Tuple[torch.Tensor, torch.Tensor]: + """Function to perform per-token-group quantization on an input tensor `x`. + It converts the tensor values into signed float8 values and returns the + quantized tensor along with the scaling factor used for quantization. + Args: + x: The input tensor with ndim >= 2. + group_size: The group size used for quantization. + eps: The minimum to avoid dividing zero. + dtype: The dype of output tensor. Note that only `torch.float8_e4m3fn` + is supported for now. + Returns: + Tuple[torch.Tensor, torch.Tensor]: The quantized tensor and the + scaling factor for quantization. + """ + if dtype is None: + dtype = ( + torch.float8_e4m3fnuz if current_platform.is_rocm() else torch.float8_e4m3fn + ) + assert x.shape[-1] % group_size == 0, ( + f"the last dimension of `x` {x.shape[-1]} must be divisible " + f"by `group_size` {group_size}" + ) + assert x.is_contiguous(), "`x` must be contiguous" + + finfo = torch.finfo(dtype) + fp8_min = finfo.min + fp8_max = finfo.max + + x_q = torch.empty_like(x, device=x.device, dtype=dtype) + M = x.numel() // group_size + N = group_size + if column_major_scales: + shape = (x.shape[-1] // group_size,) + x.shape[:-1] + x_s = torch.empty(shape, device=x.device, dtype=torch.float32).permute(-1, -2) + else: + shape = x.shape[:-1] + (x.shape[-1] // group_size,) + x_s = torch.empty(shape, device=x.device, dtype=torch.float32) + + BLOCK = triton.next_power_of_2(N) + # heuristics for number of warps + num_warps = min(max(BLOCK // 256, 1), 8) + num_stages = 1 + if column_major_scales: + _per_token_group_quant_fp8_colmajor[(M,)]( + x, + x_q, + x_s, + group_size, + x.shape[1], + x_s.stride(1), + eps, + fp8_min=fp8_min, + fp8_max=fp8_max, + BLOCK=BLOCK, + num_warps=num_warps, + num_stages=num_stages, + ) + else: + _per_token_group_quant_fp8[(M,)]( + x, + x_q, + x_s, + group_size, + eps, + fp8_min=fp8_min, + fp8_max=fp8_max, + BLOCK=BLOCK, + num_warps=num_warps, + num_stages=num_stages, + ) + + return x_q, x_s + + +@triton.jit +def _w8a8_block_fp8_matmul( + # Pointers to inputs and output + A, + B, + C, + As, + Bs, + # Shape for matmul + M, + N, + K, + # Block size for block-wise quantization + group_n, + group_k, + # Stride for inputs and output + stride_am, + stride_ak, + stride_bk, + stride_bn, + stride_cm, + stride_cn, + stride_As_m, + stride_As_k, + stride_Bs_k, + stride_Bs_n, + # Meta-parameters + BLOCK_SIZE_M: tl.constexpr, + BLOCK_SIZE_N: tl.constexpr, + BLOCK_SIZE_K: tl.constexpr, + GROUP_SIZE_M: tl.constexpr, +): + """Triton-accelerated function used to perform linear operations (dot + product) on input tensors `A` and `B` with block-wise quantization, and + store the result in output tensor `C`. + """ + + pid = tl.program_id(axis=0) + num_pid_m = tl.cdiv(M, BLOCK_SIZE_M) + num_pid_n = tl.cdiv(N, BLOCK_SIZE_N) + num_pid_in_group = GROUP_SIZE_M * num_pid_n + group_id = pid // num_pid_in_group + first_pid_m = group_id * GROUP_SIZE_M + group_size_m = min(num_pid_m - first_pid_m, GROUP_SIZE_M) + pid_m = first_pid_m + (pid % group_size_m) + pid_n = (pid % num_pid_in_group) // group_size_m + + offs_am = (pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)) % M + offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)) % N + offs_k = tl.arange(0, BLOCK_SIZE_K) + a_ptrs = A + (offs_am[:, None] * stride_am + offs_k[None, :] * stride_ak) + b_ptrs = B + (offs_k[:, None] * stride_bk + offs_bn[None, :] * stride_bn) + + As_ptrs = As + offs_am * stride_As_m + offs_bsn = offs_bn // group_n + Bs_ptrs = Bs + offs_bsn * stride_Bs_n + + accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32) + for k in range(0, tl.cdiv(K, BLOCK_SIZE_K)): + a = tl.load(a_ptrs, mask=offs_k[None, :] < K - k * BLOCK_SIZE_K, other=0.0) + b = tl.load(b_ptrs, mask=offs_k[:, None] < K - k * BLOCK_SIZE_K, other=0.0) + + k_start = k * BLOCK_SIZE_K + offs_ks = k_start // group_k + a_s = tl.load(As_ptrs + offs_ks * stride_As_k) + b_s = tl.load(Bs_ptrs + offs_ks * stride_Bs_k) + + accumulator += tl.dot(a, b) * a_s[:, None] * b_s[None, :] + a_ptrs += BLOCK_SIZE_K * stride_ak + b_ptrs += BLOCK_SIZE_K * stride_bk + + if C.dtype.element_ty == tl.bfloat16: + c = accumulator.to(tl.bfloat16) + elif C.dtype.element_ty == tl.float16: + c = accumulator.to(tl.float16) + else: + c = accumulator.to(tl.float32) + + offs_cm = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M) + offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N) + c_ptrs = C + stride_cm * offs_cm[:, None] + stride_cn * offs_cn[None, :] + c_mask = (offs_cm[:, None] < M) & (offs_cn[None, :] < N) + tl.store(c_ptrs, c, mask=c_mask) + + +@functools.lru_cache +def get_w8a8_block_fp8_configs( + N: int, K: int, block_n: int, block_k: int +) -> Optional[Dict[int, Any]]: + """ + Return optimized configurations for the w8a8 block fp8 kernel. + The return value will be a dictionary that maps an irregular grid of + batch sizes to configurations of the w8a8 block fp8 kernel. To evaluate the + kernel on a given batch size bs, the closest batch size in the grid should + be picked and the associated configuration chosen to invoke the kernel. + """ + + # First look up if an optimized configuration is available in the configs + # directory + device_name = current_platform.get_device_name().replace(" ", "_") + json_file_name = f"N={N},K={K},device_name={device_name},dtype=fp8_w8a8,block_shape=[{block_n}, {block_k}].json" # noqa: E501 + + config_file_path = os.path.join( + os.path.dirname(os.path.realpath(__file__)), "configs", json_file_name + ) + if os.path.exists(config_file_path): + with open(config_file_path) as f: + logger.info( + "Using configuration from %s for W8A8 Block FP8 kernel.", + config_file_path, + ) + # If a configuration has been found, return it + return {int(key): val for key, val in json.load(f).items()} + + # If no optimized configuration is available, we will use the default + # configuration + logger.warning( + "Using default W8A8 Block FP8 kernel config. Performance might " + "be sub-optimal! Config file not found at %s", + config_file_path, + ) + return None + + +def w8a8_block_fp8_matmul( + A: torch.Tensor, + B: torch.Tensor, + As: torch.Tensor, + Bs: torch.Tensor, + block_size: List[int], + output_dtype: torch.dtype = torch.float16, +) -> torch.Tensor: + """This function performs matrix multiplication with block-wise + quantization. + It takes two input tensors `A` and `B` with scales `As` and `Bs`. + The output is returned in the specified `output_dtype`. + Args: + A: The input tensor, e.g., activation. + B: The input tensor, e.g., weight. + As: The per-token-group quantization scale for `A`. + Bs: The per-block quantization scale for `B`. + block_size: The block size for per-block quantization. It should + be 2-dim, e.g., [128, 128]. + output_dytpe: The dtype of the returned tensor. + Returns: + torch.Tensor: The result of matmul. + """ + assert len(block_size) == 2 + block_n, block_k = block_size[0], block_size[1] + + assert A.shape[-1] == B.shape[-1] + assert A.shape[:-1] == As.shape[:-1] and A.is_contiguous() + assert triton.cdiv(A.shape[-1], block_k) == As.shape[-1] + M = A.numel() // A.shape[-1] + + assert B.ndim == 2 and B.is_contiguous() and Bs.ndim == 2 + N, K = B.shape + assert triton.cdiv(N, block_n) == Bs.shape[0] + assert triton.cdiv(K, block_k) == Bs.shape[1] + + C_shape = A.shape[:-1] + (N,) + C = A.new_empty(C_shape, dtype=output_dtype) + + configs = get_w8a8_block_fp8_configs(N, K, block_size[0], block_size[1]) + if configs: + # Get the optimal config if there is one + config = configs[min(configs.keys(), key=lambda x: abs(x - M))] + else: + # Default config + # Block-wise quant: BLOCK_SIZE_N must be divisible by block_size[0] + # BLOCK_SIZE_K must be divisible by block_size[1] + config = { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": block_size[0], + "BLOCK_SIZE_K": block_size[1], + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 2, + } + + def grid(META): + return ( + triton.cdiv(M, META["BLOCK_SIZE_M"]) * triton.cdiv(N, META["BLOCK_SIZE_N"]), + ) + + _w8a8_block_fp8_matmul[grid]( + A, + B, + C, + As, + Bs, + M, + N, + K, + block_n, + block_k, + A.stride(-2), + A.stride(-1), + B.stride(1), + B.stride(0), + C.stride(-2), + C.stride(-1), + As.stride(-2), + As.stride(-1), + Bs.stride(1), + Bs.stride(0), + **config, + ) + + return C diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/fused_marlin_moe.py b/build/torch27-cxx11-cu128-aarch64-linux/moe/fused_marlin_moe.py new file mode 100644 index 0000000000000000000000000000000000000000..b3e0a5c24599730faf973fad3cf3fb6031a30522 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/fused_marlin_moe.py @@ -0,0 +1,413 @@ +"""Fused MoE utilities for GPTQ.""" + +import functools +from typing import TYPE_CHECKING, Optional + +import torch + +from ._ops import add_op_namespace_prefix, ops +from .fused_moe import fused_topk, moe_align_block_size, try_get_optimal_moe_config +from .scalar_type import ScalarType, scalar_types + +# neuron has torch version that doesn't even have impl_abstract +if TYPE_CHECKING: + + def register_fake(fn): + return lambda name: fn + +else: + try: + from torch.library import register_fake + except ImportError: + from torch.library import impl_abstract as register_fake + + +def get_scalar_type(num_bits: int, has_zp: bool): + if has_zp: + assert num_bits == 4 + return scalar_types.uint4 + else: + return scalar_types.uint4b8 if num_bits == 4 else scalar_types.uint8b128 + + +def single_marlin_moe( + hidden_states: torch.Tensor, + w: torch.Tensor, + scales: torch.Tensor, + gating_output: torch.Tensor, + topk: int, + renormalize: bool, + g_idx: Optional[torch.Tensor] = None, + sort_indices: Optional[torch.Tensor] = None, + w_zeros: Optional[torch.Tensor] = None, + num_bits: int = 8, + is_k_full: bool = True, +) -> torch.Tensor: + """ + This function computes the multiplication of hidden_states with expert + weights used in Marlin MoE, using weights w and top-k gating mechanism. + Its purpose is testing and debugging the fused MoE kernel. + + Parameters: + - hidden_states (torch.Tensor): The input tensor to the Marlin Mul. + - w (torch.Tensor): The set of expert weights. + - scales (torch.Tensor): The quantization scales. + - gating_output (torch.Tensor): The output of the gating operation + (before softmax). + - g_idx (Optional[torch.Tensor]): Optional act_order indices. + - sort_indices (Optional[torch.Tensor]): Optional act_order input + permutation. + - topk (int): The number of top-k experts to select. + - renormalize (bool): If True, renormalize the top-k weights to sum to 1. + - w_zeros (Optional[torch.Tensor]): Optional zero points to be used for w. + - num_bits (bool): The number of bits in expert weights quantization. + + Returns: + - torch.Tensor: The output tensor after applying the MoE layer. + """ + # Check constraints. + assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch" + assert hidden_states.shape[1] == w.shape[1] * 16, "Hidden size mismatch" + assert gating_output.shape[1] == w.shape[0], "Number of experts mismatch" + assert hidden_states.is_contiguous(), "Hidden_states must be contiguous" + assert w.is_contiguous(), "Expert weights must be contiguous" + assert hidden_states.dtype == torch.float16 + assert num_bits in [4, 8] + + M, K = hidden_states.shape + E = w.shape[0] + N = w.shape[2] // (num_bits // 2) + + topk_weights, topk_ids = fused_topk(hidden_states, gating_output, topk, renormalize) + + # This might not be an optimal config for a single MMM + get_config_func = functools.partial( + try_get_optimal_moe_config, + w.shape, + w.shape, + topk_ids.shape[1], + None, + is_marlin=True, + ) + config = get_config_func(M) + + block_size_m = config["BLOCK_SIZE_M"] + + sorted_token_ids, _, _ = moe_align_block_size(topk_ids, block_size_m, E) + + max_workspace_size = (N // 64) * 16 + workspace = torch.zeros( + max_workspace_size, + dtype=torch.int, + device=hidden_states.device, + requires_grad=False, + ) + + has_zero_point = w_zeros is not None + if w_zeros is None: + w_zeros = torch.empty( + (0, 0), + dtype=hidden_states.dtype, + device=hidden_states.device, + requires_grad=False, + ) + + if g_idx is None: + g_idx = torch.empty( + (0, 0), dtype=torch.int32, device=hidden_states.device, requires_grad=False + ) + + if sort_indices is None: + sort_indices = torch.empty( + (0), dtype=torch.int32, device=hidden_states.device, requires_grad=False + ) + + scalar_type = get_scalar_type(num_bits, has_zero_point) + + intermediate_cache = ops.marlin_gemm_moe( + hidden_states, + w, + sorted_token_ids, + topk_weights, + topk_ids, + scales, + w_zeros, + g_idx, + sort_indices, + workspace, + scalar_type.id, + M, + N, + K, + is_k_full, + E, + topk, + block_size_m, + True, + False, + ) + + return torch.sum(intermediate_cache.view(*intermediate_cache.shape), dim=1) + + +if hasattr(ops, "single_marlin_gemm_moe"): + + @register_fake(add_op_namespace_prefix("single_marlin_gemm_moe")) + def single_marlin_moe_fake( + hidden_states: torch.Tensor, + w: torch.Tensor, + scales: torch.Tensor, + gating_output: torch.Tensor, + topk: int, + renormalize: bool, + g_idx: Optional[torch.Tensor] = None, + sort_indices: Optional[torch.Tensor] = None, + w_zeros: Optional[torch.Tensor] = None, + num_bits: int = 8, + is_k_full: bool = True, + ) -> torch.Tensor: + return torch.empty_like(hidden_states) + + +def fused_marlin_moe( + hidden_states: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + w1_scale: torch.Tensor, + w2_scale: torch.Tensor, + gating_output: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + g_idx1: Optional[torch.Tensor] = None, + g_idx2: Optional[torch.Tensor] = None, + sort_indices1: Optional[torch.Tensor] = None, + sort_indices2: Optional[torch.Tensor] = None, + w1_zeros: Optional[torch.Tensor] = None, + w2_zeros: Optional[torch.Tensor] = None, + num_bits: int = 8, + is_k_full: bool = True, +) -> torch.Tensor: + """ + This function computes a Mixture of Experts (MoE) layer using two sets of + weights, w1 and w2, and top-k gating mechanism. + + Parameters: + - hidden_states (torch.Tensor): The input tensor to the MoE layer. + - w1 (torch.Tensor): The first set of expert weights. + - w2 (torch.Tensor): The second set of expert weights. + - w1_scale (torch.Tensor): Scale to be used for w1. + - w2_scale (torch.Tensor): Scale to be used for w2. + - gating_output (torch.Tensor): The output of the gating operation + (before softmax). + - g_idx1 (Optional[torch.Tensor]): The first set of act_order indices. + - g_idx2 (Optional[torch.Tensor]): The second set of act_order indices. + - sort_indices1 (Optional[torch.Tensor]): The first act_order input + permutation. + - sort_indices2 (Optional[torch.Tensor]): The second act_order input + permutation. + - topk_weights (torch.Tensor): Top-k weights. + - topk_ids (torch.Tensor): Indices of topk-k elements. + - w1_zeros (Optional[torch.Tensor]): Optional zero points to be used for w1. + - w2_zeros (Optional[torch.Tensor]): Optional zero points to be used for w2. + - num_bits (bool): The number of bits in expert weights quantization. + + Returns: + - torch.Tensor: The output tensor after applying the MoE layer. + """ + # Check constraints. + assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch" + assert hidden_states.shape[1] == w1.shape[1] * 16, "Hidden size mismatch w1" + assert hidden_states.shape[1] == w2.shape[2] // ( + num_bits // 2 + ), "Hidden size mismatch w2" + assert gating_output.shape[1] == w1.shape[0], "Number of experts mismatch" + assert hidden_states.is_contiguous(), "Hidden_states must be contiguous" + assert w1.is_contiguous(), "Expert weights1 must be contiguous" + assert w2.is_contiguous(), "Expert weights2 must be contiguous" + assert hidden_states.dtype == torch.float16 + assert num_bits in [4, 8] + + has_no_act_order = ( + g_idx1 is None + and g_idx2 is None + and sort_indices1 is None + and sort_indices2 is None + ) + has_all_act_order = ( + g_idx1 is not None + and g_idx2 is not None + and sort_indices1 is not None + and sort_indices2 is not None + ) + assert has_no_act_order or has_all_act_order, ( + "g_idx and sorted_indices " "must be all not None or must be all None" + ) + + has_no_zp = w1_zeros is None and w2_zeros is None + has_all_zp = w1_zeros is not None and w2_zeros is not None + assert has_no_zp or has_all_zp, ( + "zero points must be both not None or " "must be both None" + ) + + M, K = hidden_states.shape + E = w1.shape[0] + N = w2.shape[1] * 16 + topk = topk_ids.shape[1] + + get_config_func = functools.partial( + try_get_optimal_moe_config, + w1.shape, + w2.shape, + topk_ids.shape[1], + None, + is_marlin=True, + ) + config = get_config_func(M) + + block_size_m = config["BLOCK_SIZE_M"] + + sorted_token_ids, _, _ = moe_align_block_size(topk_ids, block_size_m, E) + + max_workspace_size = (max(2 * N, K) // 64) * 16 + workspace = torch.zeros( + max_workspace_size, dtype=torch.int, device="cuda", requires_grad=False + ) + + if has_no_zp: + w1_zeros = torch.empty( + (0, 0), + dtype=hidden_states.dtype, + device=hidden_states.device, + requires_grad=False, + ) + w2_zeros = torch.empty( + (0, 0), + dtype=hidden_states.dtype, + device=hidden_states.device, + requires_grad=False, + ) + + if has_no_act_order: + g_idx1 = torch.empty( + (0, 0), dtype=torch.int32, device=hidden_states.device, requires_grad=False + ) + g_idx2 = torch.empty( + (0, 0), dtype=torch.int32, device=hidden_states.device, requires_grad=False + ) + sort_indices1 = torch.empty( + (0), dtype=torch.int32, device=hidden_states.device, requires_grad=False + ) + sort_indices2 = torch.empty( + (0, 0), dtype=torch.int32, device=hidden_states.device, requires_grad=False + ) + + scalar_type1 = get_scalar_type(num_bits, has_all_zp) + scalar_type2 = get_scalar_type(num_bits, has_all_zp) + + intermediate_cache2 = torch.empty( + (M * topk_ids.shape[1], N), + device=hidden_states.device, + dtype=hidden_states.dtype, + ) + + intermediate_cache1 = ops.marlin_gemm_moe( + hidden_states, + w1, + sorted_token_ids, + topk_weights, + topk_ids, + w1_scale, + w1_zeros, + g_idx1, + sort_indices1, + workspace, + scalar_type1.id, + M, + 2 * N, + K, + is_k_full, + E, + topk, + block_size_m, + True, + False, + ) + + ops.silu_and_mul(intermediate_cache2, intermediate_cache1.view(-1, 2 * N)) + + intermediate_cache3 = ops.marlin_gemm_moe( + intermediate_cache2, + w2, + sorted_token_ids, + topk_weights, + topk_ids, + w2_scale, + w2_zeros, + g_idx2, + sort_indices2, + workspace, + scalar_type2.id, + M, + K, + N, + is_k_full, + E, + topk, + block_size_m, + False, + True, + ) + + return torch.sum(intermediate_cache3.view(*intermediate_cache3.shape), dim=1) + + +if hasattr(ops, "fused_marlin_moe"): + + @register_fake(add_op_namespace_prefix("fused_marlin_moe")) + def fused_marlin_moe_fake( + hidden_states: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + w1_scale: torch.Tensor, + w2_scale: torch.Tensor, + gating_output: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + g_idx1: Optional[torch.Tensor] = None, + g_idx2: Optional[torch.Tensor] = None, + sort_indices1: Optional[torch.Tensor] = None, + sort_indices2: Optional[torch.Tensor] = None, + w1_zeros: Optional[torch.Tensor] = None, + w2_zeros: Optional[torch.Tensor] = None, + num_bits: int = 8, + is_k_full: bool = True, + ) -> torch.Tensor: + return torch.empty_like(hidden_states) + + +if hasattr(ops, "marlin_gemm_moe"): + + @register_fake(add_op_namespace_prefix("marlin_gemm_moe")) + def marlin_gemm_moe_fake( + a: torch.Tensor, + b_q_weights: torch.Tensor, + sorted_ids: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + b_scales: torch.Tensor, + b_zero_points: torch.Tensor, + g_idx: torch.Tensor, + perm: torch.Tensor, + workspace: torch.Tensor, + b_q_type: ScalarType, + size_m: torch.SymInt, + size_n: torch.SymInt, + size_k: torch.SymInt, + is_k_full: bool, + num_experts: int, + topk: int, + moe_block_size: int, + replicate_input: bool, + apply_weights: bool, + ) -> torch.Tensor: + return torch.empty((size_m, topk, size_n), dtype=a.dtype, device=a.device) diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/fused_moe.py b/build/torch27-cxx11-cu128-aarch64-linux/moe/fused_moe.py new file mode 100644 index 0000000000000000000000000000000000000000..f8f0586c2a58d1fb68f7d2eaed47a48f165d1e8d --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/fused_moe.py @@ -0,0 +1,1574 @@ +# SPDX-License-Identifier: Apache-2.0 +"""Fused MoE kernel.""" + +import functools +import json +import logging +import os +from typing import Any, Callable, Dict, List, Optional, Tuple + +import torch +import triton +import triton.language as tl + +from ._ops import ops +from .fp8 import scaled_fp8_quant +from .fp8_utils import per_token_group_quant_fp8 +from .moe_align_block_size import moe_align_block_size +from .platforms import current_platform +from .utils.custom_ops import direct_register_custom_op + +logger = logging.getLogger(__name__) + + +VLLM_FUSED_MOE_CHUNK_SIZE = int(os.getenv("VLLM_FUSED_MOE_CHUNK_SIZE", "32768")) +VLLM_ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON: bool = bool( + int(os.getenv("VLLM_ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON", "0")) +) + + +def cdiv(a: int, b: int) -> int: + """Ceiling division.""" + return -(a // -b) + + +def _fp8_quantize( + A: torch.Tensor, + A_scale: Optional[torch.Tensor], + block_shape: Optional[List[int]], +) -> Tuple[torch.Tensor, torch.Tensor]: + """ + Perform fp8 quantization on the inputs. If a block_shape + is provided, the output will be blocked. + """ + if block_shape is None: + A, A_scale = scaled_fp8_quant(A, A_scale) + else: + assert len(block_shape) == 2 + _, block_k = block_shape[0], block_shape[1] + A, A_scale = per_token_group_quant_fp8(A, block_k) + assert cdiv(A.shape[-1], block_k) == A_scale.shape[-1] + return A, A_scale + + +@triton.jit +def write_zeros_to_output( + c_ptr, + stride_cm, + stride_cn, + pid_n, + N, + offs_token, + token_mask, + BLOCK_SIZE_M, + BLOCK_SIZE_N, + compute_type, +): + accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=compute_type) + offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N) + c_ptrs = c_ptr + stride_cm * offs_token[:, None] + stride_cn * offs_cn[None, :] + c_mask = token_mask[:, None] & (offs_cn[None, :] < N) + tl.store(c_ptrs, accumulator, mask=c_mask) + + +@triton.jit +def fused_moe_kernel_gptq_awq( + # Pointers to matrices + a_ptr, + b_ptr, + c_ptr, + b_scale_ptr, + b_zp_ptr, + topk_weights_ptr, + sorted_token_ids_ptr, + expert_ids_ptr, + num_tokens_post_padded_ptr, + # Matrix dimensions + N: tl.constexpr, + K: tl.constexpr, + EM, + num_valid_tokens, + # The stride variables represent how much to increase the ptr by when + # moving by 1 element in a particular dimension. E.g. `stride_am` is + # how much to increase `a_ptr` by to get the element one row down + # (A has M rows). + stride_am, + stride_ak, + stride_be, + stride_bk, + stride_bn, + stride_cm, + stride_cn, + stride_bse, + stride_bsk, + stride_bsn, + stride_bze, + stride_bzk, + stride_bzn, + block_k_diviable: tl.constexpr, + group_size: tl.constexpr, + # Meta-parameters + BLOCK_SIZE_M: tl.constexpr, + BLOCK_SIZE_N: tl.constexpr, + BLOCK_SIZE_K: tl.constexpr, + GROUP_SIZE_M: tl.constexpr, + MUL_ROUTED_WEIGHT: tl.constexpr, + top_k: tl.constexpr, + compute_type: tl.constexpr, + has_zp: tl.constexpr, + use_int4_w4a16: tl.constexpr, + use_int8_w8a16: tl.constexpr, +): + """ + Implements the fused computation for a Mixture of Experts (MOE) using + token and expert matrices. + + Key Parameters: + - A: The input tensor representing tokens with shape (*, K), where '*' can + be any shape representing batches and K is the feature dimension of + each token. + - B: The stacked MOE weight tensor with shape (E, N, K), where E is + the number of experts, K is the input feature dimension, and N is + the output feature dimension. + - C: The output cache tensor with shape (M, topk, N), where M is the + total number of tokens post padding, topk is the number of times + each token is repeated, and N is the output feature dimension. + - sorted_token_ids: A tensor containing the sorted indices of tokens, + repeated topk times and arranged by the expert index they are + assigned to. + - expert_ids: A tensor containing the indices of the expert for each + block. It determines which expert matrix from B should be used for + each block in A. + This kernel performs the multiplication of a token by its corresponding + expert matrix as determined by `expert_ids`. The sorting of + `sorted_token_ids` by expert index and padding ensures divisibility by + BLOCK_SIZE_M, which is necessary to maintain consistency in block matrix + multiplication across different blocks processed by the same expert. + """ + # ----------------------------------------------------------- + # Map program ids `pid` to the block of C it should compute. + # This is done in a grouped ordering to promote L2 data reuse. + pid = tl.program_id(axis=0) + num_pid_m = tl.cdiv(EM, BLOCK_SIZE_M) + num_pid_n = tl.cdiv(N, BLOCK_SIZE_N) + num_pid_in_group = GROUP_SIZE_M * num_pid_n + group_id = pid // num_pid_in_group + first_pid_m = group_id * GROUP_SIZE_M + group_size_m = min(num_pid_m - first_pid_m, GROUP_SIZE_M) + pid_m = first_pid_m + ((pid % num_pid_in_group) % group_size_m) + pid_n = (pid % num_pid_in_group) // group_size_m + + # ---------------------------------------------------------- + # Create pointers for the first blocks of A and B. + # We will advance this pointer as we move in the K direction + # and accumulate + # `a_ptrs` is a block of [BLOCK_SIZE_M, BLOCK_SIZE_K] pointers + # `b_ptrs` is a block of [BLOCK_SIZE_K, BLOCK_SIZE_N] pointers + num_tokens_post_padded = tl.load(num_tokens_post_padded_ptr) + if pid_m * BLOCK_SIZE_M >= num_tokens_post_padded: + return + offs_token_id = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M).to(tl.int64) + offs_token = tl.load(sorted_token_ids_ptr + offs_token_id) + token_mask = offs_token < num_valid_tokens + + off_experts = tl.load(expert_ids_ptr + pid_m).to(tl.int64) + if off_experts == -1: + # ----------------------------------------------------------- + # Write back zeros to the output when the expert is not + # in the current expert parallel rank. + write_zeros_to_output( + c_ptr, + stride_cm, + stride_cn, + pid_n, + N, + offs_token, + token_mask, + BLOCK_SIZE_M, + BLOCK_SIZE_N, + compute_type, + ) + return + + offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N).to(tl.int64)) % N + offs_k = tl.arange(0, BLOCK_SIZE_K) + a_ptrs = a_ptr + ( + offs_token[:, None] // top_k * stride_am + offs_k[None, :] * stride_ak + ) + + if use_int4_w4a16: + b_ptrs = ( + b_ptr + + off_experts * stride_be + + (offs_k[:, None] // 2) * stride_bk + + offs_bn[None, :] * stride_bn + ) + b_shifter = (offs_k[:, None] % 2) * 4 + elif use_int8_w8a16: + b_ptrs = ( + b_ptr + + off_experts * stride_be + + offs_k[:, None] * stride_bk + + offs_bn[None, :] * stride_bn + ) + + if not has_zp and use_int4_w4a16: + b_zp_num = 8 + if not has_zp and use_int8_w8a16: + b_zp_num = 128 + elif has_zp and use_int4_w4a16: + b_zp_shifter = (offs_bn[None, :] % 2) * 4 + + # ----------------------------------------------------------- + # Iterate to compute a block of the C matrix. + # We accumulate into a `[BLOCK_SIZE_M, BLOCK_SIZE_N]` block + # of fp32 values for higher accuracy. + # `accumulator` will be converted back to fp16 after the loop. + accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32) + for k in range(0, tl.cdiv(K, BLOCK_SIZE_K)): + # Load the next block of A and B, generate a mask by checking the + # K dimension. + + if not block_k_diviable: + k_mask = offs_k[:, None] < K - k * BLOCK_SIZE_K + k_other = 0.0 + else: + k_mask = None + k_other = None + + a = tl.load( + a_ptrs, + mask=token_mask[:, None] & (offs_k[None, :] < K - k * BLOCK_SIZE_K), + other=0.0, + ) + b = tl.load(b_ptrs) + if use_int4_w4a16: + b = (b >> b_shifter) & 0xF + + b_scale_ptrs = ( + b_scale_ptr + + off_experts * stride_bse + + offs_bn[None, :] * stride_bsn + + ((offs_k[:, None] + BLOCK_SIZE_K * k) // group_size) * stride_bsk + ) + b_scale = tl.load(b_scale_ptrs, mask=k_mask, other=k_other) + b_scale = b_scale.to(tl.float32) + + if has_zp and use_int4_w4a16: + offs_k_true = (offs_k[:, None] + BLOCK_SIZE_K * k) // group_size + b_zp_ptrs = ( + b_zp_ptr + + off_experts * stride_bze + + (offs_bn[None, :] // 2) * stride_bzn + + offs_k_true * stride_bzk + ) + b_zp = tl.load(b_zp_ptrs, mask=k_mask, other=k_other) + b_zp = (b_zp >> b_zp_shifter) & 0xF + b_zp = b_zp.to(tl.float32) + elif has_zp and use_int8_w8a16: + offs_k_true = (offs_k[:, None] + BLOCK_SIZE_K * k) // group_size + b_zp_ptrs = ( + b_zp_ptr + + off_experts * stride_bze + + offs_bn[None, :] * stride_bzn + + offs_k_true * stride_bzk + ) + b_zp = tl.load(b_zp_ptrs, mask=k_mask, other=k_other) + b_zp = b_zp.to(tl.float32) + + # We accumulate along the K dimension. + if has_zp: + b = ((b.to(tl.float32) - b_zp) * b_scale).to(compute_type) + else: + b = ((b.to(tl.float32) - b_zp_num) * b_scale).to(compute_type) + accumulator = tl.dot(a, b, acc=accumulator) + + # Advance the ptrs to the next K block. + a_ptrs += BLOCK_SIZE_K * stride_ak + if use_int4_w4a16: + b_ptrs += (BLOCK_SIZE_K // 2) * stride_bk + else: + b_ptrs += BLOCK_SIZE_K * stride_bk + + if MUL_ROUTED_WEIGHT: + moe_weight = tl.load(topk_weights_ptr + offs_token, mask=token_mask, other=0) + accumulator = accumulator * moe_weight[:, None] + + accumulator = accumulator.to(compute_type) + # ----------------------------------------------------------- + # Write back the block of the output + offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N) + c_ptrs = c_ptr + stride_cm * offs_token[:, None] + stride_cn * offs_cn[None, :] + c_mask = token_mask[:, None] & (offs_cn[None, :] < N) + tl.store(c_ptrs, accumulator, mask=c_mask) + + +@triton.jit +def fused_moe_kernel( + # Pointers to matrices + a_ptr, + b_ptr, + c_ptr, + a_scale_ptr, + b_scale_ptr, + topk_weights_ptr, + sorted_token_ids_ptr, + expert_ids_ptr, + num_tokens_post_padded_ptr, + # Matrix dimensions + N, + K, + EM, + num_valid_tokens, + # The stride variables represent how much to increase the ptr by when + # moving by 1 element in a particular dimension. E.g. `stride_am` is + # how much to increase `a_ptr` by to get the element one row down + # (A has M rows). + stride_am, + stride_ak, + stride_be, + stride_bk, + stride_bn, + stride_cm, + stride_cn, + stride_asm, + stride_ask, + stride_bse, + stride_bsk, + stride_bsn, + # Block size for block-wise quantization + group_n: tl.constexpr, + group_k: tl.constexpr, + # Meta-parameters + BLOCK_SIZE_M: tl.constexpr, + BLOCK_SIZE_N: tl.constexpr, + BLOCK_SIZE_K: tl.constexpr, + GROUP_SIZE_M: tl.constexpr, + MUL_ROUTED_WEIGHT: tl.constexpr, + top_k: tl.constexpr, + compute_type: tl.constexpr, + use_fp8_w8a8: tl.constexpr, + use_int8_w8a16: tl.constexpr, +): + """ + Implements the fused computation for a Mixture of Experts (MOE) using + token and expert matrices. + + Key Parameters: + - A: The input tensor representing tokens with shape (*, K), where '*' can + be any shape representing batches and K is the feature dimension of + each token. + - B: The stacked MOE weight tensor with shape (E, N, K), where E is + the number of experts, K is the input feature dimension, and N is + the output feature dimension. + - C: The output cache tensor with shape (M, topk, N), where M is the + total number of tokens post padding, topk is the number of times + each token is repeated, and N is the output feature dimension. + - sorted_token_ids: A tensor containing the sorted indices of tokens, + repeated topk times and arranged by the expert index they are + assigned to. + - expert_ids: A tensor containing the indices of the expert for each + block. It determines which expert matrix from B should be used for + each block in A. + This kernel performs the multiplication of a token by its corresponding + expert matrix as determined by `expert_ids`. The sorting of + `sorted_token_ids` by expert index and padding ensures divisibility by + BLOCK_SIZE_M, which is necessary to maintain consistency in block matrix + multiplication across different blocks processed by the same expert. + """ + # ----------------------------------------------------------- + # Map program ids `pid` to the block of C it should compute. + # This is done in a grouped ordering to promote L2 data reuse. + pid = tl.program_id(axis=0) + num_pid_m = tl.cdiv(EM, BLOCK_SIZE_M) + num_pid_n = tl.cdiv(N, BLOCK_SIZE_N) + num_pid_in_group = GROUP_SIZE_M * num_pid_n + group_id = pid // num_pid_in_group + first_pid_m = group_id * GROUP_SIZE_M + group_size_m = min(num_pid_m - first_pid_m, GROUP_SIZE_M) + pid_m = first_pid_m + ((pid % num_pid_in_group) % group_size_m) + pid_n = (pid % num_pid_in_group) // group_size_m + + # ---------------------------------------------------------- + # Create pointers for the first blocks of A and B. + # We will advance this pointer as we move in the K direction + # and accumulate + # `a_ptrs` is a block of [BLOCK_SIZE_M, BLOCK_SIZE_K] pointers + # `b_ptrs` is a block of [BLOCK_SIZE_K, BLOCK_SIZE_N] pointers + num_tokens_post_padded = tl.load(num_tokens_post_padded_ptr) + if pid_m * BLOCK_SIZE_M >= num_tokens_post_padded: + return + offs_token_id = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M).to(tl.int64) + offs_token = tl.load(sorted_token_ids_ptr + offs_token_id) + token_mask = offs_token < num_valid_tokens + + off_experts = tl.load(expert_ids_ptr + pid_m).to(tl.int64) + if off_experts == -1: + # ----------------------------------------------------------- + # Write back zeros to the output when the expert is not + # in the current expert parallel rank. + write_zeros_to_output( + c_ptr, + stride_cm, + stride_cn, + pid_n, + N, + offs_token, + token_mask, + BLOCK_SIZE_M, + BLOCK_SIZE_N, + compute_type, + ) + return + + offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N).to(tl.int64)) % N + offs_k = tl.arange(0, BLOCK_SIZE_K) + a_ptrs = a_ptr + ( + offs_token[:, None] // top_k * stride_am + offs_k[None, :] * stride_ak + ) + + b_ptrs = ( + b_ptr + + off_experts * stride_be + + (offs_k[:, None] * stride_bk + offs_bn[None, :] * stride_bn) + ) + if use_int8_w8a16: + b_scale_ptrs = ( + b_scale_ptr + off_experts * stride_bse + offs_bn[None, :] * stride_bsn + ) + b_scale = tl.load(b_scale_ptrs) + + if use_fp8_w8a8: + if group_k > 0 and group_n > 0: + a_scale_ptrs = a_scale_ptr + (offs_token // top_k) * stride_asm + offs_bsn = offs_bn // group_n + b_scale_ptrs = ( + b_scale_ptr + off_experts * stride_bse + offs_bsn * stride_bsn + ) + else: + a_scale = tl.load(a_scale_ptr) + b_scale = tl.load(b_scale_ptr + off_experts) + + # ----------------------------------------------------------- + # Iterate to compute a block of the C matrix. + # We accumulate into a `[BLOCK_SIZE_M, BLOCK_SIZE_N]` block + # of fp32 values for higher accuracy. + # `accumulator` will be converted back to fp16 after the loop. + accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32) + for k in range(0, tl.cdiv(K, BLOCK_SIZE_K)): + # Load the next block of A and B, generate a mask by checking the + # K dimension. + a = tl.load( + a_ptrs, + mask=token_mask[:, None] & (offs_k[None, :] < K - k * BLOCK_SIZE_K), + other=0.0, + ) + b = tl.load(b_ptrs, mask=offs_k[:, None] < K - k * BLOCK_SIZE_K, other=0.0) + # We accumulate along the K dimension. + if use_int8_w8a16: + accumulator = tl.dot(a, b.to(compute_type), acc=accumulator) + elif use_fp8_w8a8: + if group_k > 0 and group_n > 0: + k_start = k * BLOCK_SIZE_K + offs_ks = k_start // group_k + a_scale = tl.load( + a_scale_ptrs + offs_ks * stride_ask, mask=token_mask, other=0.0 + ) + b_scale = tl.load(b_scale_ptrs + offs_ks * stride_bsk) + + accumulator += tl.dot(a, b) * a_scale[:, None] * b_scale[None, :] + else: + accumulator = tl.dot(a, b, acc=accumulator) + else: + accumulator += tl.dot(a, b) + # Advance the ptrs to the next K block. + a_ptrs += BLOCK_SIZE_K * stride_ak + b_ptrs += BLOCK_SIZE_K * stride_bk + + if MUL_ROUTED_WEIGHT: + moe_weight = tl.load(topk_weights_ptr + offs_token, mask=token_mask, other=0) + accumulator = accumulator * moe_weight[:, None] + if use_int8_w8a16: + accumulator = (accumulator * b_scale).to(compute_type) + elif use_fp8_w8a8: + if group_k > 0 and group_n > 0: + accumulator = accumulator.to(compute_type) + else: + accumulator = (accumulator * a_scale * b_scale).to(compute_type) + else: + accumulator = accumulator.to(compute_type) + # ----------------------------------------------------------- + # Write back the block of the output + offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N) + c_ptrs = c_ptr + stride_cm * offs_token[:, None] + stride_cn * offs_cn[None, :] + c_mask = token_mask[:, None] & (offs_cn[None, :] < N) + tl.store(c_ptrs, accumulator, mask=c_mask) + + +def invoke_fused_moe_kernel( + A: torch.Tensor, + B: torch.Tensor, + C: torch.Tensor, + A_scale: Optional[torch.Tensor], + B_scale: Optional[torch.Tensor], + B_zp: Optional[torch.Tensor], + topk_weights: Optional[torch.Tensor], + sorted_token_ids: torch.Tensor, + expert_ids: torch.Tensor, + num_tokens_post_padded: torch.Tensor, + mul_routed_weight: bool, + top_k: int, + config: Dict[str, Any], + compute_type: tl.dtype, + use_fp8_w8a8: bool, + use_int8_w8a16: bool, + use_int4_w4a16: bool, + block_shape: Optional[List[int]] = None, +) -> None: + assert topk_weights is not None or not mul_routed_weight + assert topk_weights is None or topk_weights.stride(1) == 1 + assert sorted_token_ids.stride(0) == 1 + + if use_fp8_w8a8: + assert B_scale is not None + assert ( + block_shape is None + or triton.cdiv(B.shape[-2], block_shape[0]) == B_scale.shape[-2] + ) + assert ( + block_shape is None + or triton.cdiv(B.shape[-1], block_shape[1]) == B_scale.shape[-1] + ) + + elif use_int8_w8a16 or use_int4_w4a16: + assert B_scale is not None + assert block_shape is None or block_shape[0] == 0 + else: + assert A_scale is None + assert B_scale is None + + M = A.shape[0] + num_tokens = M * top_k + + EM = sorted_token_ids.shape[0] + if A.shape[0] < config["BLOCK_SIZE_M"]: + # optimize for small batch_size. + # We assume that top_ids of each token is unique, so + # so num_valid_experts <= batch_size <= BLOCK_SIZE_M, + # and we can skip some invalid blocks. + EM = min(sorted_token_ids.shape[0], A.shape[0] * top_k * config["BLOCK_SIZE_M"]) + grid = lambda META: ( + triton.cdiv(EM, META["BLOCK_SIZE_M"]) + * triton.cdiv(B.shape[1], META["BLOCK_SIZE_N"]), + ) + + if ( + (use_int8_w8a16 or use_int4_w4a16) + and block_shape is not None + and block_shape[1] > 0 + ): + assert B_scale is not None and B_scale.ndim == 3 + assert B_zp is None or B_zp.ndim == 3 + + use_moe_wna16_cuda = should_moe_wna16_use_cuda( + num_valid_tokens=num_tokens, + group_size=block_shape[1], + num_experts=B.shape[0], + bit=4 if use_int4_w4a16 else 8, + ) + config = config.copy() + config.update( + get_moe_wna16_block_config( + config=config, + use_moe_wna16_cuda=use_moe_wna16_cuda, + num_valid_tokens=num_tokens, + size_k=A.shape[1], + size_n=B.shape[1], + num_experts=B.shape[1], + group_size=block_shape[1], + real_top_k=top_k, + block_size_m=config["BLOCK_SIZE_M"], + ) + ) + + if use_moe_wna16_cuda: + bit = 4 if use_int4_w4a16 else 8 + ops.moe_wna16_gemm( + A, + C, + B, + B_scale, + B_zp, + topk_weights if mul_routed_weight else None, + sorted_token_ids, + expert_ids, + num_tokens_post_padded, + top_k, + config["BLOCK_SIZE_M"], + config["BLOCK_SIZE_N"], + config["BLOCK_SIZE_K"], + bit, + ) + return + + fused_moe_kernel_gptq_awq[grid]( + A, + B, + C, + B_scale, + B_zp, + topk_weights, + sorted_token_ids, + expert_ids, + num_tokens_post_padded, + B.shape[1], + A.shape[1], + EM, + num_tokens, + A.stride(0), + A.stride(1), + B.stride(0), + B.stride(2), + B.stride(1), + C.stride(1), + C.stride(2), + B_scale.stride(0), + B_scale.stride(2), + B_scale.stride(1), + B_zp.stride(0) if B_zp is not None else 0, + B_zp.stride(2) if B_zp is not None else 0, + B_zp.stride(1) if B_zp is not None else 0, + block_k_diviable=A.shape[1] % config["BLOCK_SIZE_K"] == 0, + group_size=block_shape[1], + MUL_ROUTED_WEIGHT=mul_routed_weight, + top_k=top_k, + compute_type=compute_type, + has_zp=B_zp is not None, + use_int4_w4a16=use_int4_w4a16, + use_int8_w8a16=use_int8_w8a16, + **config, + ) + else: + config = config.copy() + BLOCK_SIZE_K = config.pop("BLOCK_SIZE_K") + if block_shape is not None: + BLOCK_SIZE_K = min(BLOCK_SIZE_K, min(block_shape[0], block_shape[1])) + fused_moe_kernel[grid]( + A, + B, + C, + A_scale, + B_scale, + topk_weights, + sorted_token_ids, + expert_ids, + num_tokens_post_padded, + B.shape[1], + B.shape[2], + EM, + num_tokens, + A.stride(0), + A.stride(1), + B.stride(0), + B.stride(2), + B.stride(1), + C.stride(1), + C.stride(2), + A_scale.stride(0) if A_scale is not None and A_scale.ndim == 2 else 0, + A_scale.stride(1) if A_scale is not None and A_scale.ndim == 2 else 0, + B_scale.stride(0) if B_scale is not None and B_scale.ndim >= 2 else 0, + B_scale.stride(2) if B_scale is not None and B_scale.ndim == 3 else 0, + B_scale.stride(1) if B_scale is not None and B_scale.ndim >= 2 else 0, + 0 if block_shape is None else block_shape[0], + 0 if block_shape is None else block_shape[1], + MUL_ROUTED_WEIGHT=mul_routed_weight, + top_k=top_k, + compute_type=compute_type, + use_fp8_w8a8=use_fp8_w8a8, + use_int8_w8a16=use_int8_w8a16, + BLOCK_SIZE_K=BLOCK_SIZE_K, + **config, + ) + + +# Adapted from: https://github.com/sgl-project/sglang/pull/2628 +def get_config_file_name( + E: int, N: int, dtype: Optional[str], block_shape: Optional[List[int]] = None +) -> str: + device_name = current_platform.get_device_name().replace(" ", "_") + dtype_selector = "" if not dtype else f",dtype={dtype}" + block_shape_selector = ( + "" if not block_shape or not all(block_shape) else f",block_shape={block_shape}" + ).replace(" ", "") + return f"E={E},N={N},device_name={device_name}{dtype_selector}{block_shape_selector}.json" # noqa: E501 + + +# Adapted from: https://github.com/sgl-project/sglang/pull/2628 +@functools.lru_cache +def get_moe_configs( + E: int, + N: int, + dtype: Optional[str], + block_n: Optional[int] = None, + block_k: Optional[int] = None, +) -> Optional[Dict[int, Any]]: + """ + Return optimized configurations for the fused MoE kernel. + + The return value will be a dictionary that maps an irregular grid of + batch sizes to configurations of the fused_moe kernel. To evaluate the + kernel on a given batch size bs, the closest batch size in the grid should + be picked and the associated configuration chosen to invoke the kernel. + """ + + # First look up if an optimized configuration is available in the configs + # directory + block_shape = [block_n, block_k] if block_n and block_k else None + json_file_name = get_config_file_name(E, N, dtype, block_shape) + + config_file_path = os.path.join( + os.path.dirname(os.path.realpath(__file__)), "configs", json_file_name + ) + if os.path.exists(config_file_path): + with open(config_file_path) as f: + logger.info("Using configuration from %s for MoE layer.", config_file_path) + # If a configuration has been found, return it + return {int(key): val for key, val in json.load(f).items()} + + # If no optimized configuration is available, we will use the default + # configuration + logger.warning( + ( + "Using default MoE config. Performance might be sub-optimal! " + "Config file not found at %s" + ), + config_file_path, + ) + return None + + +def get_moe_wna16_block_config( + config: Dict[str, int], + use_moe_wna16_cuda: bool, + num_valid_tokens: int, + size_k: int, + size_n: int, + num_experts: int, + group_size: int, + real_top_k: int, + block_size_m: int, +): + if "BLOCK_SIZE_N" in config and "BLOCK_SIZE_K" in config: + # optimal block config is set + return {} + if not use_moe_wna16_cuda: + # triton moe wna16 kernel + if num_valid_tokens // real_top_k == 1: + # if bs=1, use a smaller BLOCK_SIZE_N + return {"BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64} + else: + return {"BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32} + else: + # cuda moe wna16 kernel + # set default block_size 128, and increase them when num_blocks + # is too large. + block_size_n = 128 + block_size_k = 128 + if block_size_k <= group_size: + block_size_k = group_size + + num_n_blocks = size_k // block_size_k + num_k_blocks = size_n // block_size_k + num_m_blocks = ( + num_valid_tokens + block_size_m - 1 + ) / block_size_m + num_experts + if num_valid_tokens // real_top_k <= block_size_m: + num_m_blocks = min(num_m_blocks, num_valid_tokens) + num_blocks = num_m_blocks * num_n_blocks * num_k_blocks + + if size_k % 256 == 0 and num_blocks >= 256 and block_size_k < 256: + block_size_k = 256 + num_blocks = num_blocks // (256 // block_size_k) + + if ( + num_m_blocks <= 16 + and size_k % (block_size_k * 2) == 0 + and size_k % (block_size_k * 2) == 0 + and block_size_k <= 512 + and num_blocks >= 512 + ): + block_size_k = block_size_k * 2 + num_blocks = num_blocks // 2 + + if num_blocks > 1024: + block_size_n = 256 + num_n_blocks = num_n_blocks // 2 + num_blocks = num_blocks // 2 + + if size_n <= 1024 and num_blocks >= 1024: + # The kernel performance got much better with BLOCK_SIZE_N=1024 + # when num_blocks is large, event when N is small. + # Not sure why, maybe it force the CUDA SM process only one block + # at the same time. + block_size_n = 1024 + + return {"BLOCK_SIZE_N": block_size_n, "BLOCK_SIZE_K": block_size_k} + + +def should_moe_wna16_use_cuda( + num_valid_tokens: int, group_size: int, num_experts: int, bit: int +): + return ( + bit == 4 and group_size in [32, 64, 128] and num_valid_tokens / num_experts <= 6 + ) + + +def get_default_config( + M: int, + E: int, + N: int, + K: int, + topk: int, + dtype: Optional[str], + is_marlin: bool, + block_shape: Optional[List[int]] = None, +) -> Dict[str, int]: + if dtype == "fp8_w8a8" and block_shape is not None: + # Block-wise quant: BLOCK_SIZE_N must be divisible by block_shape[0] + # BLOCK_SIZE_K must be divisible by block_shape[1] + config = { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": block_shape[0], + "BLOCK_SIZE_K": block_shape[1], + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3, + } + elif dtype in ["int4_w4a16", "int8_w8a16"] and block_shape is not None: + # moe wna16 kernels + # only set BLOCK_SIZE_M + # BLOCK_SIZE_N and BLOCK_SIZE_K would be set later + bit = 4 if dtype == "int4_w4a16" else 8 + use_moe_wna16_cuda = should_moe_wna16_use_cuda(M * topk, block_shape[1], E, bit) + if use_moe_wna16_cuda: + config = {"BLOCK_SIZE_M": min(16, M)} + elif M <= 20: + config = {"BLOCK_SIZE_M": 16, "GROUP_SIZE_M": 1} + elif M <= 40: + config = {"BLOCK_SIZE_M": 32, "GROUP_SIZE_M": 1} + else: + config = {"BLOCK_SIZE_M": 64, "GROUP_SIZE_M": 1} + else: + config = { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 8, + } + # A heuristic: fused marlin works faster with this config for small M + if M <= E or (is_marlin and M <= 32): + config = { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + } + return config + + +def try_get_optimal_moe_config( + w1_shape: Tuple[int, ...], + w2_shape: Tuple[int, ...], + top_k: int, + dtype: Optional[str], + M: int, + is_marlin: bool = False, + block_shape: Optional[List[int]] = None, +): + # from vllm.model_executor.layers.fused_moe import get_config + # TODO: removed when syncing to vLLM, do we need this? + # override_config = get_config() + override_config = None + if override_config: + config = override_config + else: + # First try to load optimal config from the file + E, _, N = w2_shape + if dtype == "int4_w4a16": + N = N * 2 + block_n = block_shape[0] if block_shape else 0 + block_k = block_shape[1] if block_shape else 0 + configs = get_moe_configs(E, N, dtype, block_n, block_k) + + if configs: + # If an optimal configuration map has been found, look up the + # optimal config + config = configs[min(configs.keys(), key=lambda x: abs(x - M))] + else: + # Else use the default config + config = get_default_config( + M, E, N, w1_shape[2], top_k, dtype, is_marlin, block_shape + ) + return config + + +def vllm_topk_softmax( + topk_weights: torch.Tensor, + topk_indices: torch.Tensor, + token_expert_indices: torch.Tensor, + gating_output: torch.Tensor, + renormalize: bool, +) -> tuple[torch.Tensor, ...]: + ops.topk_softmax( + topk_weights, + topk_indices, + token_expert_indices, + gating_output, + ) + if renormalize: + topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True) + + return topk_weights, topk_indices + + +def dispatch_topk_func() -> Callable[..., tuple[torch.Tensor, ...]]: + return vllm_topk_softmax + + +def fused_topk( + hidden_states: torch.Tensor, + gating_output: torch.Tensor, + topk: int, + renormalize: bool, +) -> Tuple[torch.Tensor, torch.Tensor]: + assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch" + + M, _ = hidden_states.shape + + topk_weights = torch.empty( + M, topk, dtype=torch.float32, device=hidden_states.device + ) + topk_ids = torch.empty(M, topk, dtype=torch.int32, device=hidden_states.device) + token_expert_indicies = torch.empty( + M, topk, dtype=torch.int32, device=hidden_states.device + ) + + gating_output_float = gating_output.float() # TODO(woosuk): Optimize this. + + topk_func = dispatch_topk_func() + topk_weights, topk_ids = topk_func( + topk_weights, topk_ids, token_expert_indicies, gating_output_float, renormalize + ) + + del token_expert_indicies # Not used. Will be used in the future. + return topk_weights, topk_ids + + +# This is used by the Deepseek-V2 and Deepseek-V3 model +@torch.compile(dynamic=True, backend=current_platform.simple_compile_backend) +def grouped_topk( + hidden_states: torch.Tensor, + gating_output: torch.Tensor, + topk: int, + renormalize: bool, + num_expert_group: int = 0, + topk_group: int = 0, + scoring_func: str = "softmax", + e_score_correction_bias: Optional[torch.Tensor] = None, +) -> Tuple[torch.Tensor, torch.Tensor]: + + assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch" + + if scoring_func == "softmax": + scores = torch.softmax(gating_output, dim=-1) + elif scoring_func == "sigmoid": + scores = gating_output.sigmoid() + else: + raise ValueError(f"Unsupported scoring function: {scoring_func}") + + num_token = scores.shape[0] + if e_score_correction_bias is not None: + # Store original scores before applying correction bias. We use biased + # scores for expert selection but original scores for routing weights + original_scores = scores + scores = scores + e_score_correction_bias.unsqueeze(0) + group_scores = ( + scores.view(num_token, num_expert_group, -1).topk(2, dim=-1)[0].sum(dim=-1) + ) + else: + group_scores = ( + scores.view(num_token, num_expert_group, -1).max(dim=-1).values + ) # [n, n_group] + group_idx = torch.topk(group_scores, k=topk_group, dim=-1, sorted=False)[ + 1 + ] # [n, top_k_group] + group_mask = torch.zeros_like(group_scores) # [n, n_group] + group_mask.scatter_(1, group_idx, 1) # [n, n_group] + score_mask = ( + group_mask.unsqueeze(-1) + .expand(num_token, num_expert_group, scores.shape[-1] // num_expert_group) + .reshape(num_token, -1) + ) # [n, e] + tmp_scores = scores.masked_fill(~score_mask.bool(), float("-inf")) # [n, e] + + if e_score_correction_bias is not None: + topk_ids = torch.topk(tmp_scores, k=topk, dim=-1, sorted=False)[1] + # Use original unbiased scores for the routing weights + topk_weights = original_scores.gather(1, topk_ids) + else: + topk_weights, topk_ids = torch.topk(tmp_scores, k=topk, dim=-1, sorted=False) + + if renormalize: + topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True) + + return topk_weights.to(torch.float32), topk_ids.to(torch.int32) + + +def get_config_dtype_str( + dtype: torch.dtype, + use_int4_w4a16: Optional[bool] = False, + use_int8_w8a16: Optional[bool] = False, + use_fp8_w8a8: Optional[bool] = False, +) -> Optional[str]: + if use_fp8_w8a8: + return "fp8_w8a8" + elif use_int8_w8a16: + return "int8_w8a16" + elif use_int4_w4a16: + return "int4_w4a16" + elif dtype == torch.float: + # avoiding cases where kernel fails when float32 MoE + # use fp16/bfloat16 configs + return "float32" + return None + + +def inplace_fused_experts( + hidden_states: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + activation: str = "silu", + apply_router_weight_on_input: bool = False, + use_fp8_w8a8: bool = False, + use_int8_w8a16: bool = False, + use_int4_w4a16: bool = False, + global_num_experts: int = -1, + expert_map: Optional[torch.Tensor] = None, + w1_scale: Optional[torch.Tensor] = None, + w2_scale: Optional[torch.Tensor] = None, + w1_zp: Optional[torch.Tensor] = None, + w2_zp: Optional[torch.Tensor] = None, + a1_scale: Optional[torch.Tensor] = None, + a2_scale: Optional[torch.Tensor] = None, + block_shape: Optional[List[int]] = None, +) -> None: + fused_experts_impl( + hidden_states, + w1, + w2, + topk_weights, + topk_ids, + True, + activation, + apply_router_weight_on_input, + use_fp8_w8a8, + use_int8_w8a16, + use_int4_w4a16, + global_num_experts, + expert_map, + w1_scale, + w2_scale, + w1_zp, + w2_zp, + a1_scale, + a2_scale, + block_shape, + ) + + +def inplace_fused_experts_fake( + hidden_states: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + activation: str = "silu", + apply_router_weight_on_input: bool = False, + use_fp8_w8a8: bool = False, + use_int8_w8a16: bool = False, + use_int4_w4a16: bool = False, + global_num_experts: int = -1, + expert_map: Optional[torch.Tensor] = None, + w1_scale: Optional[torch.Tensor] = None, + w2_scale: Optional[torch.Tensor] = None, + w1_zp: Optional[torch.Tensor] = None, + w2_zp: Optional[torch.Tensor] = None, + a1_scale: Optional[torch.Tensor] = None, + a2_scale: Optional[torch.Tensor] = None, + block_shape: Optional[List[int]] = None, +) -> None: + pass + + +direct_register_custom_op( + op_name="inplace_fused_experts", + op_func=inplace_fused_experts, + mutates_args=["hidden_states"], + fake_impl=inplace_fused_experts_fake, +) + + +def outplace_fused_experts( + hidden_states: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + activation: str = "silu", + apply_router_weight_on_input: bool = False, + use_fp8_w8a8: bool = False, + use_int8_w8a16: bool = False, + use_int4_w4a16: bool = False, + global_num_experts: int = -1, + expert_map: Optional[torch.Tensor] = None, + w1_scale: Optional[torch.Tensor] = None, + w2_scale: Optional[torch.Tensor] = None, + w1_zp: Optional[torch.Tensor] = None, + w2_zp: Optional[torch.Tensor] = None, + a1_scale: Optional[torch.Tensor] = None, + a2_scale: Optional[torch.Tensor] = None, + block_shape: Optional[List[int]] = None, +) -> torch.Tensor: + return fused_experts_impl( + hidden_states, + w1, + w2, + topk_weights, + topk_ids, + False, + activation, + apply_router_weight_on_input, + use_fp8_w8a8, + use_int8_w8a16, + use_int4_w4a16, + global_num_experts, + expert_map, + w1_scale, + w2_scale, + w1_zp, + w2_zp, + a1_scale, + a2_scale, + block_shape, + ) + + +def outplace_fused_experts_fake( + hidden_states: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + activation: str = "silu", + apply_router_weight_on_input: bool = False, + use_fp8_w8a8: bool = False, + use_int8_w8a16: bool = False, + use_int4_w4a16: bool = False, + global_num_experts: int = -1, + expert_map: Optional[torch.Tensor] = None, + w1_scale: Optional[torch.Tensor] = None, + w2_scale: Optional[torch.Tensor] = None, + w1_zp: Optional[torch.Tensor] = None, + w2_zp: Optional[torch.Tensor] = None, + a1_scale: Optional[torch.Tensor] = None, + a2_scale: Optional[torch.Tensor] = None, + block_shape: Optional[List[int]] = None, +) -> torch.Tensor: + return torch.empty_like(hidden_states) + + +direct_register_custom_op( + op_name="outplace_fused_experts", + op_func=outplace_fused_experts, + mutates_args=[], + fake_impl=outplace_fused_experts_fake, +) + + +def torch_vllm_inplace_fused_experts(**kwargs) -> torch.Tensor: + ops.inplace_fused_experts(**kwargs) + hidden_states = kwargs["hidden_states"] + return hidden_states + + +def torch_vllm_outplace_fused_experts(**kwargs) -> torch.Tensor: + return ops.outplace_fused_experts(**kwargs) + + +def dispatch_fused_experts_func(inplace: bool) -> Callable[..., torch.Tensor]: + if inplace: + return torch_vllm_inplace_fused_experts + return torch_vllm_outplace_fused_experts + + +def fused_experts( + hidden_states: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + inplace: bool = False, + activation: str = "silu", + apply_router_weight_on_input: bool = False, + use_fp8_w8a8: bool = False, + use_int8_w8a16: bool = False, + use_int4_w4a16: bool = False, + global_num_experts: int = -1, + expert_map: Optional[torch.Tensor] = None, + w1_scale: Optional[torch.Tensor] = None, + w2_scale: Optional[torch.Tensor] = None, + w1_zp: Optional[torch.Tensor] = None, + w2_zp: Optional[torch.Tensor] = None, + a1_scale: Optional[torch.Tensor] = None, + a2_scale: Optional[torch.Tensor] = None, + block_shape: Optional[List[int]] = None, +) -> torch.Tensor: + return dispatch_fused_experts_func(inplace)( + hidden_states=hidden_states, + w1=w1, + w2=w2, + topk_weights=topk_weights, + topk_ids=topk_ids, + activation=activation, + apply_router_weight_on_input=apply_router_weight_on_input, + use_fp8_w8a8=use_fp8_w8a8, + use_int8_w8a16=use_int8_w8a16, + use_int4_w4a16=use_int4_w4a16, + global_num_experts=global_num_experts, + expert_map=expert_map, + w1_scale=w1_scale, + w2_scale=w2_scale, + w1_zp=w1_zp, + w2_zp=w2_zp, + a1_scale=a1_scale, + a2_scale=a2_scale, + block_shape=block_shape, + ) + + +def fused_experts_impl( + hidden_states: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + topk_weights: torch.Tensor, + topk_ids: torch.Tensor, + inplace: bool = False, + activation: str = "silu", + apply_router_weight_on_input: bool = False, + use_fp8_w8a8: bool = False, + use_int8_w8a16: bool = False, + use_int4_w4a16: bool = False, + global_num_experts: int = -1, + expert_map: Optional[torch.Tensor] = None, + w1_scale: Optional[torch.Tensor] = None, + w2_scale: Optional[torch.Tensor] = None, + w1_zp: Optional[torch.Tensor] = None, + w2_zp: Optional[torch.Tensor] = None, + a1_scale: Optional[torch.Tensor] = None, + a2_scale: Optional[torch.Tensor] = None, + block_shape: Optional[List[int]] = None, +): + # Check constraints. + if use_int4_w4a16: + assert hidden_states.shape[1] // 2 == w1.shape[2], "Hidden size mismatch" + else: + assert hidden_states.shape[1] == w1.shape[2], "Hidden size mismatch" + + assert topk_weights.shape == topk_ids.shape, "topk shape mismatch" + assert hidden_states.is_contiguous(), "Hidden_states must be contiguous" + assert w1.stride(-1) == 1, "Stride of last dimension must be 1" + assert w2.stride(-1) == 1, "Stride of last dimension must be 1" + assert hidden_states.dtype in [torch.float32, torch.float16, torch.bfloat16] + + num_tokens, _ = hidden_states.shape + E, N, _ = w1.shape + K = w2.shape[1] + if global_num_experts == -1: + global_num_experts = E + top_k_num = topk_ids.shape[1] + # We execute the fused_moe kernel in chunks to circumvent this issue: + # https://github.com/vllm-project/vllm/issues/5938 + CHUNK_SIZE = VLLM_FUSED_MOE_CHUNK_SIZE + M = min(num_tokens, CHUNK_SIZE) + config_dtype = get_config_dtype_str( + use_fp8_w8a8=use_fp8_w8a8, + use_int8_w8a16=use_int8_w8a16, + use_int4_w4a16=use_int4_w4a16, + dtype=hidden_states.dtype, + ) + + get_config_func = functools.partial( + try_get_optimal_moe_config, + w1.shape, + w2.shape, + top_k_num, + config_dtype, + block_shape=block_shape, + ) + + config = get_config_func(M) + + # We can reuse the memory between these because by the time we need + # cache3, we're done with cache1 + cache13 = torch.empty( + M * top_k_num * max(N, K), + device=hidden_states.device, + dtype=hidden_states.dtype, + ) + intermediate_cache1 = cache13[: M * top_k_num * N].view(M, top_k_num, N) + intermediate_cache3 = cache13[: M * top_k_num * K].view(M, top_k_num, K) + + # This needs separate memory since it's used concurrently with cache1 + intermediate_cache2 = torch.empty( + (M * top_k_num, N // 2), device=hidden_states.device, dtype=hidden_states.dtype + ) + + if hidden_states.dtype == torch.bfloat16: + compute_type = tl.bfloat16 + elif hidden_states.dtype == torch.float16: + compute_type = tl.float16 + elif hidden_states.dtype == torch.float32: + compute_type = tl.float32 + else: + raise ValueError(f"Unsupported compute_type: {hidden_states.dtype}") + + if inplace: + out_hidden_states = hidden_states + else: + out_hidden_states = torch.empty_like(hidden_states) + + for chunk in range((num_tokens // CHUNK_SIZE) + 1): + begin_chunk_idx, end_chunk_idx = ( + chunk * CHUNK_SIZE, + min((chunk + 1) * CHUNK_SIZE, num_tokens), + ) + curr_hidden_states = hidden_states[begin_chunk_idx:end_chunk_idx] + tokens_in_chunk, _ = curr_hidden_states.shape + + if tokens_in_chunk == 0: + break + + if tokens_in_chunk < CHUNK_SIZE and chunk > 0: + # Adjust the intermediate cache size and config for the last + # chunk. Note that in most cases we only have one chunk + # so the cache size and config are already set correctly and + # do not need to be adjusted. + intermediate_cache1 = intermediate_cache1[:tokens_in_chunk] + intermediate_cache2 = intermediate_cache2[ + : tokens_in_chunk * topk_ids.shape[1] + ] + intermediate_cache3 = intermediate_cache3[:tokens_in_chunk] + config = get_config_func(tokens_in_chunk) + + curr_topk_ids = topk_ids[begin_chunk_idx:end_chunk_idx] + curr_topk_weights = topk_weights[begin_chunk_idx:end_chunk_idx] + + a1q_scale: Optional[torch.Tensor] = None + + if use_fp8_w8a8: + qcurr_hidden_states, a1q_scale = _fp8_quantize( + curr_hidden_states, a1_scale, block_shape + ) + else: + qcurr_hidden_states = curr_hidden_states + a1q_scale = a1_scale + + sorted_token_ids, expert_ids, num_tokens_post_padded = moe_align_block_size( + curr_topk_ids, config["BLOCK_SIZE_M"], global_num_experts, expert_map + ) + + invoke_fused_moe_kernel( + qcurr_hidden_states, + w1, + intermediate_cache1, + a1q_scale, + w1_scale, + w1_zp, + curr_topk_weights, + sorted_token_ids, + expert_ids, + num_tokens_post_padded, + apply_router_weight_on_input, + top_k_num, + config, + compute_type=compute_type, + use_fp8_w8a8=use_fp8_w8a8, + use_int8_w8a16=use_int8_w8a16, + use_int4_w4a16=use_int4_w4a16, + block_shape=block_shape, + ) + + if activation == "silu": + ops.silu_and_mul(intermediate_cache2, intermediate_cache1.view(-1, N)) + elif activation == "gelu": + ops.gelu_and_mul(intermediate_cache2, intermediate_cache1.view(-1, N)) + else: + raise ValueError(f"Unsupported FusedMoe activation: {activation}") + + a2q_scale: Optional[torch.Tensor] = None + + if use_fp8_w8a8: + qintermediate_cache2, a2q_scale = _fp8_quantize( + intermediate_cache2, a2_scale, block_shape + ) + else: + qintermediate_cache2 = intermediate_cache2 + a2q_scale = a2_scale + + invoke_fused_moe_kernel( + qintermediate_cache2, + w2, + intermediate_cache3, + a2q_scale, + w2_scale, + w2_zp, + curr_topk_weights, + sorted_token_ids, + expert_ids, + num_tokens_post_padded, + not apply_router_weight_on_input, + 1, + config, + compute_type=compute_type, + use_fp8_w8a8=use_fp8_w8a8, + use_int8_w8a16=use_int8_w8a16, + use_int4_w4a16=use_int4_w4a16, + block_shape=block_shape, + ) + + ops.moe_sum( + intermediate_cache3.view(*intermediate_cache3.shape), + out_hidden_states[begin_chunk_idx:end_chunk_idx], + ) + + return out_hidden_states + + +def fused_moe( + hidden_states: torch.Tensor, + w1: torch.Tensor, + w2: torch.Tensor, + gating_output: torch.Tensor, + topk: int, + renormalize: bool, + inplace: bool = False, + activation: str = "silu", + apply_router_weight_on_input: bool = False, + use_grouped_topk: bool = False, + num_expert_group: Optional[int] = None, + topk_group: Optional[int] = None, + custom_routing_function: Optional[Callable] = None, + use_fp8_w8a8: bool = False, + use_int8_w8a16: bool = False, + use_int4_w4a16: bool = False, + global_num_experts: int = -1, + expert_map: Optional[torch.Tensor] = None, + w1_scale: Optional[torch.Tensor] = None, + w2_scale: Optional[torch.Tensor] = None, + w1_zp: Optional[torch.Tensor] = None, + w2_zp: Optional[torch.Tensor] = None, + a1_scale: Optional[torch.Tensor] = None, + a2_scale: Optional[torch.Tensor] = None, + block_shape: Optional[List[int]] = None, +) -> torch.Tensor: + """ + This function computes a Mixture of Experts (MoE) layer using two sets of + weights, w1 and w2, and top-k gating mechanism. + + Parameters: + - hidden_states (torch.Tensor): The input tensor to the MoE layer. + - w1 (torch.Tensor): The first set of expert weights. + - w2 (torch.Tensor): The second set of expert weights. + - gating_output (torch.Tensor): The output of the gating operation + (before softmax). + - topk (int): The number of top-k experts to select. + - renormalize (bool): If True, renormalize the top-k weights to sum to 1. + - inplace (bool): If True, perform the operation in-place. + Defaults to False. + - activation (str): The activation function to apply after the first + MoE layer. + - num_expert_group: Optional[int]: additional parameter for grouped_topk + - topk_group: Optional[int]: additional parameter for grouped_topk + - use_grouped_topk: If True, use grouped_topk instead of fused_topk + note: Deepseekv2 model uses grouped_topk + - use_fp8_w8a8 (bool): If True, use fp8 arithmetic to compute the inner + products for w1 and w2. Defaults to False. + - use_int8_w8a16 (bool): If True, use matmul of int8 weight and bf16/fp16 + activation to compute the inner products for w1 and w2. + Defaults to False. + - use_int4_w4a16 (bool): If True, use matmul of int4 weight and bf16/fp16 + activation to compute the inner products for w1 and w2. + Defaults to False. + - global_num_experts (int): The total number of experts in the global + expert space. + - expert_map (Optional[torch.Tensor]): A tensor mapping expert indices + from the global expert space to the local expert space of the expert + parallel shard. + - w1_scale (Optional[torch.Tensor]): Optional scale to be used for + w1. + - w2_scale (Optional[torch.Tensor]): Optional scale to be used for + w2. + - a1_scale (Optional[torch.Tensor]): Optional scale to be used for + a1. + - a2_scale (Optional[torch.Tensor]): Optional scale to be used for + a2. + - block_shape: (Optional[List[int]]): Optional block size for block-wise + quantization. + + Returns: + - torch.Tensor: The output tensor after applying the MoE layer. + """ + + if use_grouped_topk: + assert num_expert_group is not None and topk_group is not None + topk_weights, topk_ids = grouped_topk( + hidden_states, + gating_output, + topk, + renormalize, + num_expert_group, + topk_group, + ) + elif custom_routing_function is None: + topk_weights, topk_ids = fused_topk( + hidden_states, gating_output, topk, renormalize + ) + else: + topk_weights, topk_ids = custom_routing_function( + hidden_states, gating_output, topk, renormalize + ) + + return fused_experts( + hidden_states, + w1, + w2, + topk_weights, + topk_ids, + inplace=inplace, + activation=activation, + apply_router_weight_on_input=apply_router_weight_on_input, + use_fp8_w8a8=use_fp8_w8a8, + use_int8_w8a16=use_int8_w8a16, + use_int4_w4a16=use_int4_w4a16, + global_num_experts=global_num_experts, + expert_map=expert_map, + w1_scale=w1_scale, + w2_scale=w2_scale, + w1_zp=w1_zp, + w2_zp=w2_zp, + a1_scale=a1_scale, + a2_scale=a2_scale, + block_shape=block_shape, + ) diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/layers.py b/build/torch27-cxx11-cu128-aarch64-linux/moe/layers.py new file mode 100644 index 0000000000000000000000000000000000000000..1befea3f9e97a6a4aa1679821d30091c3d825bb8 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/layers.py @@ -0,0 +1,64 @@ +import torch +import torch.nn as nn + +from .fused_moe import fused_moe + + +def _llama4_topk( + hidden_states: torch.Tensor, + gating_output: torch.Tensor, + topk: int, + renormalize: bool, +): + assert hidden_states.shape[0] == gating_output.shape[0], "Number of tokens mismatch" + topk_weights, topk_ids = torch.topk(gating_output, topk, dim=1) + topk_weights = torch.sigmoid(topk_weights.float()).to(hidden_states.dtype) + return topk_weights, topk_ids + + +def _fix_llama4_experts(hidden_states: torch.Tensor, experts: nn.Module): + if experts.gate_up_proj.shape[-1] != hidden_states.shape[-1]: + experts.gate_up_proj = nn.Parameter( + experts.gate_up_proj.transpose(1, 2).contiguous() + ) + experts.down_proj = nn.Parameter(experts.down_proj.transpose(1, 2).contiguous()) + + +class Llama4TextMoe(nn.Module): + has_backward = False + + experts: nn.Module + router: nn.Linear + shared_expert: nn.Module + top_k: int + + def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: + hidden_states = hidden_states.reshape(-1, hidden_states.shape[-1]) + + _fix_llama4_experts(hidden_states, self.experts) + + router_logits = self.router(hidden_states) + + extra_kwargs = {} + use_fp8_w8a8 = False + if hasattr(self.experts, "gate_up_proj_scale"): + use_fp8_w8a8 = True + extra_kwargs["w1_scale"] = self.experts.gate_up_proj_scale + extra_kwargs["w2_scale"] = self.experts.down_proj_scale + + out = fused_moe( + hidden_states, + w1=self.experts.gate_up_proj, + w2=self.experts.down_proj, + gating_output=router_logits, + topk=self.top_k, + renormalize=False, + custom_routing_function=_llama4_topk, + apply_router_weight_on_input=True, + use_fp8_w8a8=use_fp8_w8a8, + **extra_kwargs + ) + + out += self.shared_expert(hidden_states) + + return out, router_logits.t() diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/moe_align_block_size.py b/build/torch27-cxx11-cu128-aarch64-linux/moe/moe_align_block_size.py new file mode 100644 index 0000000000000000000000000000000000000000..a698de5aaa123df5a2d663f252e1d5914e880d09 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/moe_align_block_size.py @@ -0,0 +1,252 @@ +# SPDX-License-Identifier: Apache-2.0 +import os +from typing import Optional, Tuple + +import torch +import triton +import triton.language as tl + +from ._ops import ops + + +def round_up(x: int, y: int) -> int: + return ((x + y - 1) // y) * y + + +VLLM_ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON = bool( + int(os.getenv("VLLM_ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON", "0")) +) + + +def ceil_div(a, b): + return (a + b - 1) // b + + +@triton.jit +def moe_align_block_size_stage1( + topk_ids_ptr, + tokens_cnts_ptr, + num_experts: tl.constexpr, + numel: tl.constexpr, + tokens_per_thread: tl.constexpr, +): + pid = tl.program_id(0) + + start_idx = pid * tokens_per_thread + + off_c = (pid + 1) * num_experts + + for i in range(tokens_per_thread): + if start_idx + i < numel: + idx = tl.load(topk_ids_ptr + start_idx + i) + token_cnt = tl.load(tokens_cnts_ptr + off_c + idx) + tl.store(tokens_cnts_ptr + off_c + idx, token_cnt + 1) + + +@triton.jit +def moe_align_block_size_stage2( + tokens_cnts_ptr, + num_experts: tl.constexpr, +): + pid = tl.program_id(0) + + last_cnt = 0 + for i in range(1, num_experts + 1): + token_cnt = tl.load(tokens_cnts_ptr + i * num_experts + pid) + last_cnt = last_cnt + token_cnt + tl.store(tokens_cnts_ptr + i * num_experts + pid, last_cnt) + + +@triton.jit +def moe_align_block_size_stage3( + total_tokens_post_pad_ptr, + tokens_cnts_ptr, + cumsum_ptr, + num_experts: tl.constexpr, + block_size: tl.constexpr, +): + last_cumsum = 0 + off_cnt = num_experts * num_experts + for i in range(1, num_experts + 1): + token_cnt = tl.load(tokens_cnts_ptr + off_cnt + i - 1) + last_cumsum = last_cumsum + tl.cdiv(token_cnt, block_size) * block_size + tl.store(cumsum_ptr + i, last_cumsum) + tl.store(total_tokens_post_pad_ptr, last_cumsum) + + +@triton.jit +def moe_align_block_size_stage4( + topk_ids_ptr, + sorted_token_ids_ptr, + expert_ids_ptr, + tokens_cnts_ptr, + cumsum_ptr, + num_experts: tl.constexpr, + block_size: tl.constexpr, + numel: tl.constexpr, + tokens_per_thread: tl.constexpr, +): + pid = tl.program_id(0) + start_idx = tl.load(cumsum_ptr + pid) + end_idx = tl.load(cumsum_ptr + pid + 1) + + for i in range(start_idx, end_idx, block_size): + tl.store(expert_ids_ptr + i // block_size, pid) + + start_idx = pid * tokens_per_thread + off_t = pid * num_experts + + for i in range(start_idx, tl.minimum(start_idx + tokens_per_thread, numel)): + expert_id = tl.load(topk_ids_ptr + i) + token_cnt = tl.load(tokens_cnts_ptr + off_t + expert_id) + rank_post_pad = token_cnt + tl.load(cumsum_ptr + expert_id) + tl.store(sorted_token_ids_ptr + rank_post_pad, i) + tl.store(tokens_cnts_ptr + off_t + expert_id, token_cnt + 1) + + +# Triton implementation based on: +# https://github.com/sgl-project/sglang/commit/ba5112ff691d791a9e38c6c71f59324a5fcb49d0 +def moe_align_block_size_triton( + topk_ids: torch.Tensor, + num_experts: int, + block_size: int, + sorted_token_ids: torch.Tensor, + expert_ids: torch.Tensor, + num_tokens_post_pad: torch.Tensor, +) -> None: + numel = topk_ids.numel() + grid = (num_experts,) + tokens_cnts = torch.zeros( + (num_experts + 1, num_experts), dtype=torch.int32, device=topk_ids.device + ) + cumsum = torch.zeros((num_experts + 1,), dtype=torch.int32, device=topk_ids.device) + tokens_per_thread = ceil_div(numel, num_experts) + + moe_align_block_size_stage1[grid]( + topk_ids, + tokens_cnts, + num_experts, + numel, + tokens_per_thread, + ) + moe_align_block_size_stage2[grid]( + tokens_cnts, + num_experts, + ) + moe_align_block_size_stage3[(1,)]( + num_tokens_post_pad, + tokens_cnts, + cumsum, + num_experts, + block_size, + ) + moe_align_block_size_stage4[grid]( + topk_ids, + sorted_token_ids, + expert_ids, + tokens_cnts, + cumsum, + num_experts, + block_size, + numel, + tokens_per_thread, + ) + + +def moe_align_block_size( + topk_ids: torch.Tensor, + block_size: int, + num_experts: int, + expert_map: Optional[torch.Tensor] = None, + pad_sorted_ids: bool = False, +) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + """ + Aligns the token distribution across experts to be compatible with block + size for matrix multiplication. + + Parameters: + - topk_ids: A tensor of shape [total_tokens, top_k] representing the + top-k expert indices for each token. + - block_size: The block size used in block matrix multiplication. + - num_experts: The total number of experts. + - expert_map: A tensor of shape [num_experts] that maps the expert index + from the global space to the local index space of the current + expert parallel shard. If the expert is not in the current expert + parallel shard, the mapping is set to -1. + - pad_sorted_ids: A flag indicating whether the sorted_token_ids length + should be padded to a multiple of block_size, + + Returns: + - sorted_token_ids: A tensor containing the sorted token indices according + to their allocated expert. + - expert_ids: A tensor indicating the assigned expert index for each block. + - num_tokens_post_padded: The total number of tokens after padding, + ensuring divisibility by block_size. + + This function pads the number of tokens that each expert needs to process + so that it is divisible by block_size. + Padding ensures that during block matrix multiplication, the dimensions + align correctly. + + Example: + Given topk_ids = [[2, 3, 4], [1, 2, 4], [1, 3, 4], [1, 2, 3]], + block_size = 4, and num_experts = 4: + - We initially have 12 tokens (after repeating 'top_k' times) and 4 experts, + with each expert needing to process 3 tokens. + - As block_size is 4, we pad 1 token for each expert. + - First, flatten topk_ids to [2, 3, 4, 1, 2, 4, 1, 3, 4, 1, 2, 3]. + - Then append padding tokens [12, 12, 12, 12] for each block. + - After sorting by expert index, we obtain token_ids + [3, 6, 9, 12, 0, 4, 10, 12, 1, 7, 11, 12, 2, 5, 8, 12]. + Tokens 12 are non-existent (padding) and are ignored in + the subsequent matrix multiplication. + - The padding ensures that the total number of tokens is now divisible + by block_size for proper block matrix operations. + """ + max_num_tokens_padded = topk_ids.numel() + num_experts * (block_size - 1) + if pad_sorted_ids: + max_num_tokens_padded = round_up(max_num_tokens_padded, block_size) + sorted_ids = torch.empty( + (max_num_tokens_padded,), dtype=torch.int32, device=topk_ids.device + ) + sorted_ids.fill_(topk_ids.numel()) + max_num_m_blocks = triton.cdiv(max_num_tokens_padded, block_size) + # Expert ids must be zeroed out to prevent index out of bounds error while + # mapping global expert ids to local expert ids in expert parallelism. + expert_ids = torch.zeros( + (max_num_m_blocks,), dtype=torch.int32, device=topk_ids.device + ) + num_tokens_post_pad = torch.empty((1), dtype=torch.int32, device=topk_ids.device) + if num_experts >= 224: + if VLLM_ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON or num_experts != 256: + moe_align_block_size_triton( + topk_ids, + num_experts, + block_size, + sorted_ids, + expert_ids, + num_tokens_post_pad, + ) + else: + # Currently requires num_experts=256 + ops.sgl_moe_align_block_size( + topk_ids, + num_experts, + block_size, + sorted_ids, + expert_ids, + num_tokens_post_pad, + ) + else: + ops.moe_align_block_size( + topk_ids, + num_experts, + block_size, + sorted_ids, + expert_ids, + num_tokens_post_pad, + ) + if expert_map is not None: + expert_ids = expert_map[expert_ids] + + return sorted_ids, expert_ids, num_tokens_post_pad diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/platforms.py b/build/torch27-cxx11-cu128-aarch64-linux/moe/platforms.py new file mode 100644 index 0000000000000000000000000000000000000000..db83d0f2b0fadb15c14e8887ac27e4a07b6c4d62 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/platforms.py @@ -0,0 +1,99 @@ +import random +from abc import ABC, abstractmethod +from functools import lru_cache +from typing import NamedTuple, Optional + +import torch + +IS_ROCM = torch.version.hip is not None + + +class DeviceCapability(NamedTuple): + major: int + minor: int + + def as_version_str(self) -> str: + return f"{self.major}.{self.minor}" + + def to_int(self) -> int: + """ + Express device capability as an integer ````. + + It is assumed that the minor version is always a single digit. + """ + assert 0 <= self.minor < 10 + return self.major * 10 + self.minor + + +class Platform(ABC): + simple_compile_backend: str = "inductor" + + @classmethod + @abstractmethod + def get_device_name(cls, device_id: int = 0) -> str: ... + + @abstractmethod + def is_cuda_alike(self) -> bool: + """Stateless version of :func:`torch.cuda.is_available`.""" + ... + + @abstractmethod + def is_rocm(self): ... + + @classmethod + def seed_everything(cls, seed: Optional[int] = None) -> None: + """ + Set the seed of each random module. + `torch.manual_seed` will set seed on all devices. + + Loosely based on: https://github.com/Lightning-AI/pytorch-lightning/blob/2.4.0/src/lightning/fabric/utilities/seed.py#L20 + """ + if seed is not None: + random.seed(seed) + torch.manual_seed(seed) + try: + import numpy as np + + np.random.seed(seed) + except ImportError: + pass + + +class CudaPlatform(Platform): + @classmethod + def get_device_capability(cls, device_id: int = 0) -> DeviceCapability: + major, minor = torch.cuda.get_device_capability(device_id) + return DeviceCapability(major=major, minor=minor) + + @classmethod + @lru_cache(maxsize=8) + def get_device_name(cls, device_id: int = 0) -> str: + return torch.cuda.get_device_name(0) + + def is_cuda_alike(self) -> bool: + return True + + def is_rocm(self): + return False + + +class RocmPlatform(Platform): + @classmethod + @lru_cache(maxsize=8) + def get_device_capability(cls, device_id: int = 0) -> DeviceCapability: + major, minor = torch.cuda.get_device_capability(device_id) + return DeviceCapability(major=major, minor=minor) + + @classmethod + @lru_cache(maxsize=8) + def get_device_name(cls, device_id: int = 0) -> str: + return torch.cuda.get_device_name(device_id) + + def is_cuda_alike(self) -> bool: + return True + + def is_rocm(self): + return True + + +current_platform = RocmPlatform() if IS_ROCM else CudaPlatform() diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/scalar_type.py b/build/torch27-cxx11-cu128-aarch64-linux/moe/scalar_type.py new file mode 100644 index 0000000000000000000000000000000000000000..ea749fe8247b6846620ccbba30ddf48d914ca4e1 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/scalar_type.py @@ -0,0 +1,336 @@ +import functools +import struct +from dataclasses import dataclass +from enum import Enum +from typing import Optional, Union + + +# Mirrors enum in `core/scalar_type.hpp` +class NanRepr(Enum): + NONE = 0 # nans are not supported + IEEE_754 = 1 # nans are: Exp all 1s, mantissa not all 0s + EXTD_RANGE_MAX_MIN = 2 # nans are: Exp all 1s, mantissa all 1s + + +# This ScalarType class is a parallel implementation of the C++ ScalarType +# class found in csrc/core/scalar_type.hpp. These two classes should be kept +# in sync until the inductor fully supports custom C++ classes. +@dataclass(frozen=True) +class ScalarType: + """ + ScalarType can represent a wide range of floating point and integer + types, in particular it can be used to represent sub-byte data types + (something that torch.dtype currently does not support). It is also + capable of representing types with a bias, i.e.: + `stored_value = value + bias`, + this is useful for quantized types (e.g. standard GPTQ 4bit uses a bias + of 8). The implementation for this class can be found in + csrc/core/scalar_type.hpp, these type signatures should be kept in sync + with that file. + """ + + exponent: int + """ + Number of bits in the exponent if this is a floating point type + (zero if this an integer type) + """ + + mantissa: int + """ + Number of bits in the mantissa if this is a floating point type, + or the number bits representing an integer excluding the sign bit if + this an integer type. + """ + + signed: bool + "If the type is signed (i.e. has a sign bit)" + + bias: int + """ + bias used to encode the values in this scalar type + (value = stored_value - bias, default 0) for example if we store the + type as an unsigned integer with a bias of 128 then the value 0 will be + stored as 128 and -1 will be stored as 127 and 1 will be stored as 129. + """ + + _finite_values_only: bool = False + """ + Private: if infs are supported, used `has_infs()` instead. + """ + + nan_repr: NanRepr = NanRepr.IEEE_754 + """ + How NaNs are represent in this scalar type, returns NanRepr value. + (not applicable for integer types) + """ + + def _floating_point_max_int(self) -> int: + assert ( + self.mantissa <= 52 and self.exponent <= 11 + ), f"Cannot represent max/min as a double for type {self.__str__()}" + + max_mantissa = (1 << self.mantissa) - 1 + if self.nan_repr == NanRepr.EXTD_RANGE_MAX_MIN: + max_mantissa = max_mantissa - 1 + + max_exponent = (1 << self.exponent) - 2 + if self.nan_repr == NanRepr.EXTD_RANGE_MAX_MIN or self.nan_repr == NanRepr.NONE: + assert ( + self.exponent < 11 + ), f"Cannot represent max/min as a double for type {self.__str__()}" + max_exponent = max_exponent + 1 + + # adjust the exponent to match that of a double + # for now we assume the exponent bias is the standard 2^(e-1) -1, (where + # e is the exponent bits), there is some precedent for non-standard + # biases, example `float8_e4m3b11fnuz` here: + # https://github.com/jax-ml/ml_dtypes but to avoid premature over + # complication we are just assuming the standard exponent bias until + # there is a need to support non-standard biases + exponent_bias = (1 << (self.exponent - 1)) - 1 + exponent_bias_double = (1 << 10) - 1 # double e = 11 + + max_exponent_double = max_exponent - exponent_bias + exponent_bias_double + + # shift the mantissa and exponent into the proper positions for an + # IEEE double and bitwise-or them together. + return (max_mantissa << (52 - self.mantissa)) | (max_exponent_double << 52) + + def _floating_point_max(self) -> float: + double_raw = self._floating_point_max_int() + return struct.unpack("!d", struct.pack("!Q", double_raw))[0] + + def _raw_max(self) -> Union[int, float]: + if self.is_floating_point(): + return self._floating_point_max() + else: + assert ( + self.size_bits < 64 or self.size_bits == 64 and self.is_signed() + ), "Cannot represent max as an int" + return (1 << self.mantissa) - 1 + + def _raw_min(self) -> Union[int, float]: + if self.is_floating_point(): + assert ( + self.is_signed() + ), "We currently assume all floating point types are signed" + sign_bit_double = 1 << 63 + + max_raw = self._floating_point_max_int() + min_raw = max_raw | sign_bit_double + return struct.unpack("!d", struct.pack("!Q", min_raw))[0] + else: + assert ( + not self.is_signed() or self.size_bits <= 64 + ), "Cannot represent min as a int64_t" + + if self.is_signed(): + return -(1 << (self.size_bits - 1)) + else: + return 0 + + @functools.cached_property + def id(self) -> int: + """ + Convert the ScalarType to an int which can be passed to pytorch custom + ops. This layout of the int must be kept in sync with the C++ + ScalarType's from_id method. + """ + val = 0 + offset = 0 + + def or_and_advance(member, bit_width): + nonlocal val + nonlocal offset + bit_mask = (1 << bit_width) - 1 + val = val | (int(member) & bit_mask) << offset + offset = offset + bit_width + + or_and_advance(self.exponent, 8) + or_and_advance(self.mantissa, 8) + or_and_advance(self.signed, 1) + or_and_advance(self.bias, 32) + or_and_advance(self._finite_values_only, 1) + or_and_advance(self.nan_repr.value, 8) + + assert offset <= 64, f"ScalarType fields too big {offset} to fit into an int64" + + return val + + @property + def size_bits(self) -> int: + return self.exponent + self.mantissa + int(self.signed) + + def min(self) -> Union[int, float]: + """ + Min representable value for this scalar type. + (accounting for bias if there is one) + """ + return self._raw_min() - self.bias + + def max(self) -> Union[int, float]: + """ + Max representable value for this scalar type. + (accounting for bias if there is one) + """ + return self._raw_max() - self.bias + + def is_signed(self) -> bool: + """ + If the type is signed (i.e. has a sign bit), same as `signed` + added for consistency with: + https://pytorch.org/docs/stable/generated/torch.Tensor.is_signed.html + """ + return self.signed + + def is_floating_point(self) -> bool: + "If the type is a floating point type" + return self.exponent != 0 + + def is_integer(self) -> bool: + "If the type is an integer type" + return self.exponent == 0 + + def has_bias(self) -> bool: + "If the type has a non-zero bias" + return self.bias != 0 + + def has_infs(self) -> bool: + "If the type is floating point and supports infinity" + return not self._finite_values_only + + def has_nans(self) -> bool: + return self.nan_repr != NanRepr.NONE.value + + def is_ieee_754(self) -> bool: + """ + If the type is a floating point type that follows IEEE 754 + conventions + """ + return self.nan_repr == NanRepr.IEEE_754.value and not self._finite_values_only + + def __str__(self) -> str: + """ + naming generally follows: https://github.com/jax-ml/ml_dtypes + for floating point types (leading f) the scheme is: + `float_em[flags]` + flags: + - no-flags: means it follows IEEE 754 conventions + - f: means finite values only (no infinities) + - n: means nans are supported (non-standard encoding) + for integer types the scheme is: + `[u]int[b]` + - if bias is not present it means its zero + """ + if self.is_floating_point(): + ret = ( + "float" + + str(self.size_bits) + + "_e" + + str(self.exponent) + + "m" + + str(self.mantissa) + ) + + if not self.is_ieee_754(): + if self._finite_values_only: + ret = ret + "f" + if self.nan_repr != NanRepr.NONE: + ret = ret + "n" + + return ret + else: + ret = ("int" if self.is_signed() else "uint") + str(self.size_bits) + if self.has_bias(): + ret = ret + "b" + str(self.bias) + return ret + + def __repr__(self) -> str: + return "ScalarType." + self.__str__() + + # __len__ needs to be defined (and has to throw TypeError) for pytorch's + # opcheck to work. + def __len__(self) -> int: + raise TypeError + + # + # Convenience Constructors + # + + @classmethod + def int_(cls, size_bits: int, bias: Optional[int]) -> "ScalarType": + "Create a signed integer scalar type (size_bits includes sign-bit)." + ret = cls(0, size_bits - 1, True, bias if bias else 0) + ret.id # noqa B018: make sure the id is cached + return ret + + @classmethod + def uint(cls, size_bits: int, bias: Optional[int]) -> "ScalarType": + """Create a unsigned integer scalar type.""" + ret = cls(0, size_bits, False, bias if bias else 0) + ret.id # noqa B018: make sure the id is cached + return ret + + @classmethod + def float_IEEE754(cls, exponent: int, mantissa: int) -> "ScalarType": + """ + Create a standard floating point type + (i.e. follows IEEE 754 conventions). + """ + assert mantissa > 0 and exponent > 0 + ret = cls(exponent, mantissa, True, 0) + ret.id # noqa B018: make sure the id is cached + return ret + + @classmethod + def float_( + cls, exponent: int, mantissa: int, finite_values_only: bool, nan_repr: NanRepr + ) -> "ScalarType": + """ + Create a non-standard floating point type + (i.e. does not follow IEEE 754 conventions). + """ + assert mantissa > 0 and exponent > 0 + assert nan_repr != NanRepr.IEEE_754, ( + "use `float_IEEE754` constructor for floating point types that " + "follow IEEE 754 conventions" + ) + ret = cls(exponent, mantissa, True, 0, finite_values_only, nan_repr) + ret.id # noqa B018: make sure the id is cached + return ret + + +# naming generally follows: https://github.com/jax-ml/ml_dtypes +# for floating point types (leading f) the scheme is: +# `float_em[flags]` +# flags: +# - no-flags: means it follows IEEE 754 conventions +# - f: means finite values only (no infinities) +# - n: means nans are supported (non-standard encoding) +# for integer types the scheme is: +# `[u]int[b]` +# - if bias is not present it means its zero + + +class scalar_types: + int4 = ScalarType.int_(4, None) + uint4 = ScalarType.uint(4, None) + int8 = ScalarType.int_(8, None) + uint8 = ScalarType.uint(8, None) + float8_e4m3fn = ScalarType.float_(4, 3, True, NanRepr.EXTD_RANGE_MAX_MIN) + float8_e5m2 = ScalarType.float_IEEE754(5, 2) + float16_e8m7 = ScalarType.float_IEEE754(8, 7) + float16_e5m10 = ScalarType.float_IEEE754(5, 10) + + # fp6, https://github.com/usyd-fsalab/fp6_llm/tree/main + float6_e3m2f = ScalarType.float_(3, 2, True, NanRepr.NONE) + + # "gptq" types + uint2b2 = ScalarType.uint(2, 2) + uint3b4 = ScalarType.uint(3, 4) + uint4b8 = ScalarType.uint(4, 8) + uint8b128 = ScalarType.uint(8, 128) + + # colloquial names + bfloat16 = float16_e8m7 + float16 = float16_e5m10 diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/utils/__init__.py b/build/torch27-cxx11-cu128-aarch64-linux/moe/utils/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/utils/custom_ops.py b/build/torch27-cxx11-cu128-aarch64-linux/moe/utils/custom_ops.py new file mode 100644 index 0000000000000000000000000000000000000000..26ef7c23d1bea7a162d81746ac041baefb472f46 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/utils/custom_ops.py @@ -0,0 +1,68 @@ +from typing import Callable, Optional + +import torch +from torch.library import Library + +from .._ops import ops +from ..platforms import current_platform + + +# Some backends use pytorch version < 2.4.0 which doesn't +# support `torch.library.custom_op`. +def supports_custom_op() -> bool: + return hasattr(torch.library, "custom_op") + + +# TODO: we probably need a shorthand for the ops name. +vllm_lib = Library(ops.__name__.split(".")[-1], "FRAGMENT") + + +def direct_register_custom_op( + op_name: str, + op_func: Callable, + mutates_args: list[str], + fake_impl: Optional[Callable] = None, + target_lib: Optional[Library] = None, + dispatch_key: str = "CUDA", +): + """ + `torch.library.custom_op` can have significant overhead because it + needs to consider complicated dispatching logic. This function + directly registers a custom op and dispatches it to the CUDA backend. + See https://gist.github.com/youkaichao/ecbea9ec9fc79a45d2adce1784d7a9a5 + for more details. + + By default, the custom op is registered to the vLLM library. If you + want to register it to a different library, you can pass the library + object to the `target_lib` argument. + + IMPORTANT: the lifetime of the operator is tied to the lifetime of the + library object. If you want to bind the operator to a different library, + make sure the library object is alive when the operator is used. + """ + if not supports_custom_op(): + from ..platforms import current_platform + + assert not current_platform.is_cuda_alike(), ( + "cuda platform needs torch>=2.4 to support custom op, " + "chances are you are using an old version of pytorch " + "or a custom build of pytorch. It is recommended to " + "use vLLM in a fresh new environment and let it install " + "the required dependencies." + ) + return + + import torch.library + + if hasattr(torch.library, "infer_schema"): + schema_str = torch.library.infer_schema(op_func, mutates_args=mutates_args) + else: + # for pytorch 2.4 + import torch._custom_op.impl + + schema_str = torch._custom_op.impl.infer_schema(op_func, mutates_args) + my_lib = target_lib or vllm_lib + my_lib.define(op_name + schema_str) + my_lib.impl(op_name, op_func, dispatch_key=dispatch_key) + if fake_impl is not None: + my_lib._register_fake(op_name, fake_impl) diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/utils/marlin_utils.py b/build/torch27-cxx11-cu128-aarch64-linux/moe/utils/marlin_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..f27b71af67a6f5f18c8dba271926e3d40d518879 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/utils/marlin_utils.py @@ -0,0 +1,306 @@ +from typing import List, Optional, Tuple + +import numpy +import torch + +from ..scalar_type import ScalarType, scalar_types +from .quant_utils import pack_cols, unpack_cols + +GPTQ_MARLIN_TILE = 16 +GPTQ_MARLIN_MIN_THREAD_N = 64 +GPTQ_MARLIN_MIN_THREAD_K = 128 +GPTQ_MARLIN_MAX_PARALLEL = 16 + +GPTQ_MARLIN_24_TILE = 16 +GPTQ_MARLIN_24_MIN_THREAD_N = 128 +GPTQ_MARLIN_24_MIN_THREAD_K = 128 +GPTQ_MARLIN_24_MAX_PARALLEL = 64 + +GPTQ_MARLIN_24_SUPPORTED_QUANT_TYPES = [scalar_types.uint4b8, scalar_types.uint8b128] +GPTQ_MARLIN_24_SUPPORTED_GROUP_SIZES = [-1, 128] + +MARLIN_QQQ_TILE = 16 +MARLIN_QQQ_MIN_THREAD_N = 64 +MARLIN_QQQ_MIN_THREAD_K = 128 +MARLIN_QQQ_MAX_PARALLEL = 16 + +MARLIN_QQQ_SUPPORTED_NUM_BITS = [4] +MARLIN_QQQ_SUPPORTED_GROUP_SIZES = [-1, 128] +MARLIN_QQQ_SUPPORTED_SYM = [True] + +MARLIN_SUPPORTED_GROUP_SIZES = [-1, 32, 64, 128] + +# In case there is a performance issue with Marlin, the variable below can be +# changed to False, which allows Marlin to perform global reductions in fp16 +# precision (instead of fp32), and therefore, save on some memory movements. +USE_FP32_REDUCE_DEFAULT = True + + +# For binary size and compile time, we don't support the same types for with and +# without runtime zero-point. We support common cases, i.e. AWQ and GPTQ. +# TODO: we may want to move this into the C++ so its closer to the actual impl +def query_marlin_supported_quant_types( + has_zp: bool, device_capability: Optional[int] = None +): + if device_capability is None: + capability_tuple = torch.cuda.get_device_capability() + device_capability = capability_tuple[0] * 10 + capability_tuple[1] + + if device_capability < 80: + return [] + + if has_zp: + # AWQ style, unsigned + runtime zero-point + return [scalar_types.uint4, scalar_types.uint8] + else: + # GPTQ style, unsigned + symmetric bias + # TODO: once fp8_marlin is merged into "gptq_marlin" we should be able + # to add `scalar_types.float8_e4m3fn` here + return [scalar_types.uint4b8, scalar_types.uint8b128] + + +def _check_marlin_supported( + quant_type: ScalarType, + group_size: Optional[int], + has_zp: bool, + device_capability: Optional[int] = None, +) -> Tuple[bool, Optional[str]]: + + if device_capability is None: + capability_tuple = torch.cuda.get_device_capability() + device_capability = capability_tuple[0] * 10 + capability_tuple[1] + + supported_types = query_marlin_supported_quant_types(has_zp, device_capability) + + if quant_type not in supported_types: + return ( + False, + f"Marlin does not support weight_bits = {quant_type}. " + f"Only types = {supported_types} " + f"are supported (for group_size = {group_size}, " + f"device_capability = {device_capability}, zp = {has_zp}).", + ) + if group_size is None or group_size not in MARLIN_SUPPORTED_GROUP_SIZES: + return ( + False, + f"Marlin does not support group_size = {group_size}. " + f"Only group_sizes = {MARLIN_SUPPORTED_GROUP_SIZES} " + "are supported.", + ) + + return True, None + + +def check_marlin_supported( + quant_type: ScalarType, + group_size: int, + has_zp: bool = False, + device_capability: Optional[int] = None, +) -> bool: + cond, _ = _check_marlin_supported(quant_type, group_size, has_zp, device_capability) + return cond + + +def verify_marlin_supported( + quant_type: ScalarType, group_size: int, has_zp: bool = False +) -> None: + cond, err_msg = _check_marlin_supported(quant_type, group_size, has_zp) + if not cond: + assert err_msg is not None + raise ValueError(err_msg) + + +def verify_marlin_supports_shape( + output_size_per_partition: int, + input_size_per_partition: int, + input_size: int, + group_size: int, +) -> None: + + # Validate output_size_per_partition + if output_size_per_partition % GPTQ_MARLIN_MIN_THREAD_N != 0: + raise ValueError( + f"Weight output_size_per_partition = " + f"{output_size_per_partition} is not divisible by " + f" min_thread_n = {GPTQ_MARLIN_MIN_THREAD_N}. " + "Consider reducing tensor_parallel_size or running " + "with --quantization gptq." + ) + + # Validate input_size_per_partition + if input_size_per_partition % GPTQ_MARLIN_MIN_THREAD_K != 0: + raise ValueError( + f"Weight input_size_per_partition = " + f"{input_size_per_partition} is not divisible " + f"by min_thread_k = {GPTQ_MARLIN_MIN_THREAD_K}. " + "Consider reducing tensor_parallel_size or running " + "with --quantization gptq." + ) + + if group_size < input_size and input_size_per_partition % group_size != 0: + raise ValueError( + f"Weight input_size_per_partition = {input_size_per_partition}" + f" is not divisible by group_size = {group_size}." + "Consider reducing tensor_parallel_size or running " + "with --quantization gptq." + ) + + +def check_marlin_supports_shape( + output_size_per_partition: int, + input_size_per_partition: int, + input_size: int, + group_size: int, +) -> Tuple[bool, Optional[str]]: + try: + verify_marlin_supports_shape( + output_size_per_partition, input_size_per_partition, input_size, group_size + ) + except ValueError as e: + return False, e.__str__() + return True, None + + +def marlin_make_workspace( + output_size_per_partition: int, device: torch.device +) -> torch.Tensor: + max_workspace_size = ( + output_size_per_partition // GPTQ_MARLIN_MIN_THREAD_N + ) * GPTQ_MARLIN_MAX_PARALLEL + + return torch.zeros( + max_workspace_size, dtype=torch.int, device=device, requires_grad=False + ) + + +def marlin_is_k_full(act_order: bool, is_row_parallel: bool) -> bool: + return (not act_order) or (act_order and not is_row_parallel) + + +def marlin_repeat_scales_on_all_ranks( + act_order: bool, group_size: int, is_row_parallel: bool +) -> bool: + # Need to repeat scales on every rank if act_ordering or + # channelwise and RowParallelLinear + is_channelwise = group_size == -1 + return act_order or (is_channelwise and is_row_parallel) + + +def marlin_make_empty_g_idx(device: torch.device) -> torch.Tensor: + return torch.nn.Parameter( + torch.empty(0, dtype=torch.int, device=device), requires_grad=False + ) + + +def marlin_make_empty_zp(device: torch.device) -> torch.Tensor: + return torch.nn.Parameter( + torch.empty(0, dtype=torch.int, device=device), requires_grad=False + ) + + +def marlin_sort_g_idx(g_idx: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]: + g_idx_sort_indices = torch.argsort(g_idx).to(torch.int) + return g_idx[g_idx_sort_indices], g_idx_sort_indices + + +def get_scale_perms(): + scale_perm: List[int] = [] + for i in range(8): + scale_perm.extend([i + 8 * j for j in range(8)]) + scale_perm_single: List[int] = [] + for i in range(4): + scale_perm_single.extend([2 * i + j for j in [0, 1, 8, 9, 16, 17, 24, 25]]) + return scale_perm, scale_perm_single + + +def marlin_permute_scales( + s: torch.Tensor, size_k: int, size_n: int, group_size: int +) -> torch.Tensor: + + scale_perm, scale_perm_single = get_scale_perms() + if group_size < size_k and group_size != -1: + s = s.reshape((-1, len(scale_perm)))[:, scale_perm] + else: + s = s.reshape((-1, len(scale_perm_single)))[:, scale_perm_single] + s = s.reshape((-1, size_n)).contiguous() + + return s + + +def marlin_moe_permute_scales( + s: torch.Tensor, + size_k: int, + size_n: int, + group_size: int, +): + num_experts = s.shape[0] + output = torch.empty( + (num_experts, s.shape[1], s.shape[2]), + device=s.device, + dtype=s.dtype, + ) + + for e in range(num_experts): + output[e] = marlin_permute_scales(s[e], size_k, size_n, group_size) + return output + + +def marlin_zero_points( + zp: torch.Tensor, size_k: int, size_n: int, num_bits: int +) -> torch.Tensor: + # Permute zero-points in a similar way to scales, but do not use the + # "single" permutation, since zero-points are applied on every MMA + scale_perm, _ = get_scale_perms() + zp = zp.reshape((-1, len(scale_perm)))[:, scale_perm] + + # Interleave column dim (for the dequantize code) and pack it to int32 + if num_bits == 4: + interleave = numpy.array([0, 2, 4, 6, 1, 3, 5, 7]) + elif num_bits == 8: + interleave = numpy.array([0, 2, 1, 3]) + else: + raise Exception("num_bits must be 4 or 8, got {}".format(num_bits)) + + zp = zp.reshape((-1, len(interleave)))[:, interleave].ravel() + zp = zp.reshape((-1, size_n)).contiguous() + zp = pack_cols(zp, num_bits, size_k, size_n) + + return zp + + +def awq_to_marlin_zero_points( + q_zp_packed: torch.Tensor, size_k: int, size_n: int, num_bits: int +) -> torch.Tensor: + # AWQ zero-points are quantized and packed on the column dim. + # In addition, the values are permuted based on dequantizer. + # Here we undo both of these, and then apply marlin permutation + # and pack it back. + q_zp = unpack_cols(q_zp_packed, num_bits, size_k, size_n) + + # Undo interleaving (use argsort(..) to get inverse perm) + if num_bits == 4: + undo_interleave = numpy.argsort(numpy.array([0, 2, 4, 6, 1, 3, 5, 7])) + elif num_bits == 8: + undo_interleave = numpy.argsort(numpy.array([0, 2, 1, 3])) + else: + raise Exception("num_bits must be 4 or 8, got {}".format(num_bits)) + + q_zp = q_zp.reshape((-1, len(undo_interleave)))[:, undo_interleave].ravel() + q_zp = q_zp.reshape((-1, size_n)).contiguous() + + marlin_zp = marlin_zero_points(q_zp, size_k, size_n, num_bits) + return marlin_zp + + +def moe_awq_to_marlin_zero_points( + q_zp_packed: torch.Tensor, size_k: int, size_n: int, num_bits: int +): + num_experts = q_zp_packed.shape[0] + output = torch.empty( + (num_experts, q_zp_packed.shape[1], q_zp_packed.shape[2]), + device=q_zp_packed.device, + dtype=q_zp_packed.dtype, + ) + for e in range(num_experts): + output[e] = awq_to_marlin_zero_points(q_zp_packed[e], size_k, size_n, num_bits) + return output diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/utils/marlin_utils_test.py b/build/torch27-cxx11-cu128-aarch64-linux/moe/utils/marlin_utils_test.py new file mode 100644 index 0000000000000000000000000000000000000000..b03555f5bffa8d48187d37e53f17e6e3138472cc --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/utils/marlin_utils_test.py @@ -0,0 +1,161 @@ +"""Utility functions used for tests and benchmarks""" + +from typing import List, Optional + +import numpy as np +import torch + +from ..scalar_type import ScalarType +from .marlin_utils import GPTQ_MARLIN_TILE, marlin_permute_scales, marlin_zero_points +from .quant_utils import ( + get_pack_factor, + gptq_quantize_weights, + quantize_weights, + sort_weights, +) + + +class MarlinWorkspace: + + def __init__(self, out_features, min_thread_n, max_parallel): + assert ( + out_features % min_thread_n == 0 + ), "out_features = {} is undivisible by min_thread_n = {}".format( + out_features, min_thread_n + ) + + max_workspace_size = (out_features // min_thread_n) * max_parallel + + self.scratch = torch.zeros(max_workspace_size, dtype=torch.int, device="cuda") + + +def marlin_permute_weights(q_w, size_k, size_n, perm, tile=GPTQ_MARLIN_TILE): + assert q_w.shape == (size_k, size_n) + assert size_k % tile == 0, f"size_k = {size_k}, tile = {tile}" + assert size_n % tile == 0, f"size_k = {size_n}, tile = {tile}" + + # Permute weights to 16x64 marlin tiles + q_w = q_w.reshape((size_k // tile, tile, size_n // tile, tile)) + q_w = q_w.permute((0, 2, 1, 3)) + q_w = q_w.reshape((size_k // tile, size_n * tile)) + + q_w = q_w.reshape((-1, perm.numel()))[:, perm].reshape(q_w.shape) + + return q_w + + +def marlin_weights(q_w, size_k, size_n, num_bits, perm): + # Permute + q_w = marlin_permute_weights(q_w, size_k, size_n, perm) + + # Pack + pack_factor = get_pack_factor(num_bits) + orig_device = q_w.device + + q_w = q_w.cpu().numpy().astype(np.uint32) + + q_packed = np.zeros((q_w.shape[0], q_w.shape[1] // pack_factor), dtype=np.uint32) + for i in range(pack_factor): + q_packed |= q_w[:, i::pack_factor] << num_bits * i + + q_packed = torch.from_numpy(q_packed.astype(np.int32)).to(orig_device) + + return q_packed + + +def get_weight_perm(num_bits: int): + perm_list: List[int] = [] + for i in range(32): + perm1: List[int] = [] + col = i // 4 + for block in [0, 1]: + for row in [ + 2 * (i % 4), + 2 * (i % 4) + 1, + 2 * (i % 4 + 4), + 2 * (i % 4 + 4) + 1, + ]: + perm1.append(16 * row + col + 8 * block) + for j in range(4): + perm_list.extend([p + 256 * j for p in perm1]) + + perm = np.array(perm_list) + + if num_bits == 4: + interleave = np.array([0, 2, 4, 6, 1, 3, 5, 7]) + elif num_bits == 8: + interleave = np.array([0, 2, 1, 3]) + else: + raise Exception("num_bits must be 4 or 8, got {}".format(num_bits)) + + perm = perm.reshape((-1, len(interleave)))[:, interleave].ravel() + perm = torch.from_numpy(perm) + return perm + + +def marlin_quantize( + w: torch.Tensor, + quant_type: ScalarType, + group_size: int, + act_order: bool, + test_perm: Optional[torch.Tensor] = None, +): + size_k, size_n = w.shape + num_bits = quant_type.size_bits + + # Normalize group_size + if group_size == -1: + group_size = size_k + assert group_size <= size_k + + # Quantize (and apply act_order if provided) + w_ref, q_w, s, g_idx, rand_perm = gptq_quantize_weights( + w, quant_type, group_size, act_order, test_perm + ) + + # For act_order, sort the "weights" and "g_idx" so that group ids are + # increasing + sort_indices = torch.empty(0, dtype=torch.int, device=w.device) + if act_order: + q_w, g_idx, sort_indices = sort_weights(q_w, g_idx) + + # Reformat to marlin + weight_perm = get_weight_perm(num_bits) + marlin_q_w = marlin_weights(q_w, size_k, size_n, num_bits, weight_perm) + marlin_s = marlin_permute_scales(s, size_k, size_n, group_size) + + # Create result + res_list = [w_ref, marlin_q_w, marlin_s, g_idx, sort_indices, rand_perm] + for i in range(len(res_list)): + res_list[i] = res_list[i].to(w.device) + + return res_list + + +def awq_marlin_quantize(w: torch.Tensor, quant_type: ScalarType, group_size: int): + size_k, size_n = w.shape + + # Normalize group_size + if group_size == -1: + group_size = size_k + assert group_size <= size_k + + # Detect num groups + assert size_k % group_size == 0 + num_groups = size_k // group_size + + # Quantize with zp + w_ref, q_w, s, zp = quantize_weights(w, quant_type, group_size, zero_points=True) + + # Reformat to marlin + weight_perm = get_weight_perm(quant_type.size_bits) + marlin_q_w = marlin_weights(q_w, size_k, size_n, quant_type.size_bits, weight_perm) + marlin_s = marlin_permute_scales(s, size_k, size_n, group_size) + marlin_zp = marlin_zero_points(zp, num_groups, size_n, quant_type.size_bits) + + # Create result + res_list = [w_ref, marlin_q_w, marlin_s, marlin_zp] + for i in range(len(res_list)): + res_list[i] = res_list[i].to(w.device) + + return res_list diff --git a/build/torch27-cxx11-cu128-aarch64-linux/moe/utils/quant_utils.py b/build/torch27-cxx11-cu128-aarch64-linux/moe/utils/quant_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..65c90821773b60cca2b0f9102ca7ad4e3cd009d5 --- /dev/null +++ b/build/torch27-cxx11-cu128-aarch64-linux/moe/utils/quant_utils.py @@ -0,0 +1,470 @@ +"""This file is used for /tests and /benchmarks""" + +from typing import List, Optional + +import numpy +import torch + +from ..scalar_type import ScalarType, scalar_types + +SUPPORTED_GPTQ_QUANT_TYPES = [scalar_types.uint4b8, scalar_types.uint8b128] +SUPPORTED_GROUP_SIZES = [-1, 32, 64, 128] + +MARLIN_QQQ_SUPPORTED_NUM_BITS = [4] + +# Note: this is a hack. We should update each model to register the +# stacked params and get it from there instead in a future PR. +# fused_name: List[shard_name] +FUSED_LAYER_NAME_MAPPING = { + "qkv_proj": ["q_proj", "k_proj", "v_proj"], + "gate_up_proj": ["gate_proj", "up_proj"], +} + + +def pack_quantized_values_into_int32( + w_q: torch.Tensor, wtype: ScalarType, packed_dim: int = 0 +): + # move dim to pack to the end + perm = (*[i for i in range(len(w_q.shape)) if i != packed_dim], packed_dim) + inv_perm = tuple(perm.index(i) for i in range(len(perm))) + w_q_perm = w_q.permute(perm) + + pack_factor = 32 // wtype.size_bits + mask = (1 << wtype.size_bits) - 1 + + new_shape_perm = list(w_q_perm.shape) + assert w_q_perm.shape[-1] % pack_factor == 0 + new_shape_perm[-1] //= pack_factor + + res = torch.zeros(new_shape_perm, dtype=torch.int32, device=w_q.device) + for i in range(pack_factor): + res |= (w_q_perm[..., i::pack_factor] & mask) << wtype.size_bits * i + + return res.permute(inv_perm) + + +def unpack_quantized_values_into_int32( + w_q: torch.Tensor, wtype: ScalarType, packed_dim: int = 0 +): + # move dim to pack to the end + perm = (*[i for i in range(len(w_q.shape)) if i != packed_dim], packed_dim) + inv_perm = tuple(perm.index(i) for i in range(len(perm))) + w_q_perm = w_q.permute(perm) + + pack_factor = 32 // wtype.size_bits + mask = (1 << wtype.size_bits) - 1 + + new_shape_perm = list(w_q_perm.shape) + new_shape_perm[-1] *= pack_factor + + res = torch.zeros(new_shape_perm, dtype=torch.int32, device=w_q.device) + for i in range(pack_factor): + res[..., i::pack_factor] = (w_q_perm >> wtype.size_bits * i) & mask + + return res.permute(inv_perm) + + +def is_layer_skipped(prefix: str, ignored_layers: List[str]) -> bool: + # prefix: model.layers.0.self_attn.q_proj + # proj_name: q_proj + proj_name = prefix.split(".")[-1] + if proj_name in FUSED_LAYER_NAME_MAPPING: + shard_prefixes = [ + prefix.replace(proj_name, shard_proj_name) + for shard_proj_name in FUSED_LAYER_NAME_MAPPING[proj_name] + ] + + is_skipped = None + for shard_prefix in shard_prefixes: + is_shard_skipped = shard_prefix in ignored_layers + + if is_skipped is None: + is_skipped = is_shard_skipped + elif is_shard_skipped != is_skipped: + raise ValueError( + f"Detected some but not all shards of {prefix} " + "are quantized. All shards of fused layers " + "to have the same precision." + ) + else: + is_skipped = prefix in ignored_layers + + assert is_skipped is not None + return is_skipped + + +def get_pack_factor(num_bits): + assert 32 % num_bits == 0, f"Unsupported num_bits = {num_bits}" + return 32 // num_bits + + +def permute_rows( + q_w: torch.Tensor, + w_ref: torch.Tensor, + group_size: int, + test_perm: Optional[torch.Tensor] = None, +): + assert q_w.shape == w_ref.shape + + orig_device = q_w.device + k_size, _ = q_w.shape + + g_idx = torch.zeros((k_size,), dtype=torch.int32) + for i in range(k_size): + g_idx[i] = i // group_size + + # Simulate act_order by doing a random permutation on K + rand_perm = test_perm if test_perm is not None else torch.randperm(k_size) + + g_idx = g_idx[rand_perm].contiguous() + q_w = q_w[rand_perm, :].contiguous() + w_ref = w_ref[rand_perm, :].contiguous() + + return ( + w_ref.to(device=orig_device), + q_w.to(device=orig_device), + g_idx.to(device=orig_device), + rand_perm.to(device=orig_device), + ) + + +def quantize_weights( + w: torch.Tensor, + quant_type: ScalarType, + group_size: Optional[int], + zero_points: bool = False, + ref_zero_points_after_scales: bool = False, +): + assert ( + quant_type.is_integer() + ), "Floating point quantization may work but has not been tested" + assert not zero_points or group_size is not None, ( + "to have group zero points, group_size must be provided " + "(-1 group_size is channelwise)" + ) + + orig_device = w.device + orig_type = w.dtype + size_k, size_n = w.shape + + assert w.is_floating_point(), "w must be float" + + if group_size == -1: + group_size = size_k + + # Reshape to [groupsize, -1] + if group_size is not None and group_size < size_k: + w = w.reshape((-1, group_size, size_n)) + w = w.permute(1, 0, 2) + w = w.reshape((group_size, -1)) + + # Compute scale for each group + max_val = torch.max(w, 0, keepdim=True).values + min_val = torch.min(w, 0, keepdim=True).values + + max_q_val = quant_type.max() + min_q_val = quant_type.min() + + w_s = torch.Tensor([1.0]).to(w.device) # unscaled case + maybe_w_zp = None + if group_size is not None: + if zero_points: + assert not quant_type.is_signed() and quant_type.max() > 0 + w_s = (max_val - min_val).clamp(min=1e-5) / quant_type.max() + maybe_w_zp = ( + torch.round(torch.abs(min_val / w_s)).clamp(min_q_val, max_q_val).int() + ) + else: + # If the bias is such that there are no possible negative/positive + # values, set the max value to inf to avoid divide by 0 + w_s = torch.max( + abs(max_val / (max_q_val if max_q_val != 0 else torch.inf)), + abs(min_val / (min_q_val if min_q_val != 0 else torch.inf)), + ) + + # Quantize + w_q = torch.round(w / w_s).int() + (maybe_w_zp if zero_points else 0) + w_q = torch.clamp(w_q, min_q_val, max_q_val) + + # Compute ref (dequantized) + # For some kernels (namely Machete) the zero-points are applied after the + # scales are applied, for this case computing the reference in similar way + # allows us to use tighter error tolerances in our unit tests. + if ref_zero_points_after_scales and maybe_w_zp is not None: + w_ref = w_q.to(orig_type) * w_s - maybe_w_zp.to(orig_type) * w_s + else: + w_ref = (w_q - (maybe_w_zp if zero_points else 0)).to(orig_type) * w_s + + if quant_type.has_bias(): + w_q += quant_type.bias + + # Restore original shapes + if group_size is not None and group_size < size_k: + + def reshape_w(w): + w = w.reshape((group_size, -1, size_n)) + w = w.permute(1, 0, 2) + w = w.reshape((size_k, size_n)).contiguous() + return w + + w_q = reshape_w(w_q) + w_ref = reshape_w(w_ref) + w_s = w_s.reshape((-1, size_n)).contiguous() + + if maybe_w_zp is not None: + maybe_w_zp = maybe_w_zp.reshape((-1, size_n)).contiguous() + maybe_w_zp = maybe_w_zp.to(device=orig_device) + + return ( + w_ref.to(device=orig_device), + w_q.to(device=orig_device), + w_s if group_size is not None else None, + maybe_w_zp, + ) + + +def gptq_quantize_weights( + w: torch.Tensor, + quant_type: ScalarType, + group_size: int, + act_order: bool, + test_perm: Optional[torch.Tensor] = None, +): + size_k, _ = w.shape + + assert w.is_floating_point(), "w must be float" + assert ( + quant_type in SUPPORTED_GPTQ_QUANT_TYPES + ), f"Unsupported gptq type = {quant_type}" + assert group_size in SUPPORTED_GROUP_SIZES + [ + size_k + ], f"Unsupported groupsize = {group_size}" + + w_ref, w_q, w_s, _ = quantize_weights(w, quant_type, group_size) + + # Apply act_order + g_idx = torch.empty(0, dtype=torch.int, device=w.device) + rand_perm = torch.empty(0, dtype=torch.int, device=w.device) + if act_order: + assert ( + group_size < size_k + ), "For act_order, groupsize = {} must be less than size_k = {}".format( + group_size, size_k + ) + + w_ref, w_q, g_idx, rand_perm = permute_rows(w_q, w_ref, group_size, test_perm) + + return w_ref, w_q, w_s, g_idx, rand_perm + + +# QQQ employs different quant schemes for per-group and +# per-channel quantization. +def qqq_quantize_weights(w: torch.Tensor, num_bits: int, group_size: int): + orig_device = w.device + size_k, size_n = w.shape + + assert w.is_floating_point(), "w must be float" + assert ( + num_bits in MARLIN_QQQ_SUPPORTED_NUM_BITS + ), f"Unsupported num_bits = {num_bits}" + assert group_size in SUPPORTED_GROUP_SIZES + [ + size_k + ], f"Unsupported groupsize = {group_size}" + + if group_size == -1: + group_size = size_k + assert group_size <= size_k + + if group_size < size_k: + # Reshape to [groupsize, -1] + w = w.reshape((-1, group_size, size_n)) + w = w.permute(1, 0, 2) + w = w.reshape((group_size, -1)) + + max_q_val = 2**num_bits - 1 + half_q_val = (max_q_val + 1) // 2 + + # Compute scale for each group + s_group = torch.max(torch.abs(w), 0, keepdim=True)[0] + s_group *= 2 / max_q_val # 2 => symmetric + + # Quantize + q_w = torch.round(w / s_group).int() + q_w += half_q_val + q_w = torch.clamp(q_w, 0, max_q_val) + # Compute ref (dequantized) + w_ref = (q_w - half_q_val).half() * s_group + + # Restore original shapes + def reshape_w(w): + w = w.reshape((group_size, -1, size_n)) + w = w.permute(1, 0, 2) + w = w.reshape((size_k, size_n)).contiguous() + return w + + q_w = reshape_w(q_w) + w_ref = reshape_w(w_ref) + + # Compute int8 quantization scale for each channel + s_channel = torch.max(torch.abs(w_ref), 0, keepdim=True)[0] + s_channel /= 127.0 + t_int8 = (w_ref / s_channel).round().clamp(-128, 127).to(torch.int8) + w_ref = t_int8.half() * s_channel + s_channel = s_channel.reshape(1, -1).to(dtype=torch.float) + + # Fuse scales + s_group = (s_group.reshape(-1, size_n).contiguous() / s_channel).to( + dtype=torch.half + ) + else: + max_q_val = 2 ** (num_bits - 1) - 1 + + # Compute scale for each channel + s_channel = torch.max(torch.abs(w), 0, keepdim=True)[0] + s_channel /= max_q_val + + # Quantize + q_w = torch.round(w / s_channel).int() + q_w = torch.clamp(q_w, -max_q_val, max_q_val) + # Compute ref (dequantized) + w_ref = q_w.half() * s_channel + + s_group = torch.tensor([], dtype=torch.half) + # div 2 ** (8 - self.bits)) to offset right shift in unpacking + s_channel /= 2 ** (8 - num_bits) + s_channel = s_channel.reshape(-1, size_n).contiguous().to(torch.float) + + return ( + w_ref.to(device=orig_device), + q_w.to(device=orig_device), + s_group.to(device=orig_device), + s_channel.to(device=orig_device), + ) + + +def sort_weights(q_w: torch.Tensor, g_idx: torch.Tensor): + orig_device = q_w.device + + sort_indices = torch.argsort(g_idx).to(dtype=torch.int32) # Sort based on g_idx + + g_idx = g_idx[sort_indices].contiguous() + q_w = q_w[sort_indices, :].contiguous() + + return ( + q_w.to(device=orig_device), + g_idx.to(device=orig_device), + sort_indices.to(device=orig_device), + ) + + +def pack_rows( + q_w: torch.Tensor, + num_bits: int, + size_k: int, + size_n: int, +): + assert q_w.shape == (size_k, size_n) + + pack_factor = get_pack_factor(num_bits) + assert size_k % pack_factor == 0 + + orig_device = q_w.device + + q_w = q_w.cpu().numpy().astype(numpy.uint32) + + q_res = numpy.zeros((size_k // pack_factor, size_n), dtype=numpy.uint32) + + for i in range(pack_factor): + q_res |= q_w[i::pack_factor, :] << num_bits * i + + q_res = torch.from_numpy(q_res.astype(numpy.int32)).to(orig_device) + return q_res + + +def pack_cols( + q_w: torch.Tensor, + num_bits: int, + size_k: int, + size_n: int, +): + assert q_w.shape == (size_k, size_n) + + pack_factor = get_pack_factor(num_bits) + assert size_n % pack_factor == 0 + + orig_device = q_w.device + + q_w = q_w.cpu().numpy().astype(numpy.uint32) + + q_res = numpy.zeros((size_k, size_n // pack_factor), dtype=numpy.uint32) + + for i in range(pack_factor): + q_res |= q_w[:, i::pack_factor] << num_bits * i + + q_res = torch.from_numpy(q_res.astype(numpy.int32)).to(orig_device) + q_res = q_res.contiguous() + + return q_res + + +def unpack_cols( + packed_q_w: torch.Tensor, + num_bits: int, + size_k: int, + size_n: int, +): + pack_factor = get_pack_factor(num_bits) + assert size_n % pack_factor == 0 + assert packed_q_w.shape == ( + size_k, + size_n // pack_factor, + ), "packed_q_w.shape = {} size_k = {}, size_n = {} pack_Factor = {}".format( + packed_q_w.shape, size_k, size_n, pack_factor + ) + + orig_device = packed_q_w.device + + packed_q_w_cpu = packed_q_w.cpu().numpy().astype(numpy.uint32) + q_res = numpy.zeros((size_k, size_n), dtype=numpy.uint32) + + mask = (1 << num_bits) - 1 + for i in range(pack_factor): + vals = packed_q_w_cpu & mask + packed_q_w_cpu >>= num_bits + q_res[:, i::pack_factor] = vals + + q_res = torch.from_numpy(q_res.astype(numpy.int32)).to(orig_device) + q_res = q_res.contiguous() + + return q_res + + +def gptq_pack( + q_w: torch.Tensor, + num_bits: int, + size_k: int, + size_n: int, +): + return pack_rows(q_w, num_bits, size_k, size_n) + + +def awq_pack( + q_w: torch.Tensor, + num_bits: int, + size_k: int, + size_n: int, +): + assert q_w.shape == (size_k, size_n) + + # Interleave column dim (for the dequantize code) and pack it to int32 + if num_bits == 4: + interleave = numpy.array([0, 2, 4, 6, 1, 3, 5, 7]) + elif num_bits == 8: + interleave = numpy.array([0, 2, 1, 3]) + else: + raise Exception("num_bits must be 4 or 8, got {}".format(num_bits)) + + q_w = q_w.reshape((-1, len(interleave)))[:, interleave].ravel() + q_w = q_w.reshape((-1, size_n)).contiguous() + + return pack_cols(q_w, num_bits, size_k, size_n)