Export ops at the top-level
Browse files- ext-torch/moe/__init__.py +135 -1
- ext-torch/moe/_custom_ops.py +0 -135
- ext-torch/moe/fused_marlin_moe.py +1 -1
- ext-torch/moe/fused_moe.py +1 -1
ext-torch/moe/__init__.py
CHANGED
@@ -1 +1,135 @@
|
|
1 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import TYPE_CHECKING
|
2 |
+
|
3 |
+
import torch
|
4 |
+
|
5 |
+
# neuron has torch version that doesn't even have impl_abstract
|
6 |
+
if TYPE_CHECKING:
|
7 |
+
|
8 |
+
def register_fake(fn):
|
9 |
+
return lambda name: fn
|
10 |
+
|
11 |
+
else:
|
12 |
+
try:
|
13 |
+
from torch.library import register_fake
|
14 |
+
except ImportError:
|
15 |
+
from torch.library import impl_abstract as register_fake
|
16 |
+
|
17 |
+
from ._ops import add_op_namespace_prefix, ops
|
18 |
+
from .fused_marlin_moe import fused_marlin_moe
|
19 |
+
from .fused_moe import fused_moe, fused_topk, grouped_topk
|
20 |
+
from .scalar_type import ScalarType, scalar_types
|
21 |
+
|
22 |
+
|
23 |
+
def gptq_marlin_moe_repack(
|
24 |
+
b_q_weight: torch.Tensor,
|
25 |
+
perm: torch.Tensor,
|
26 |
+
size_k: int,
|
27 |
+
size_n: int,
|
28 |
+
num_bits: int,
|
29 |
+
) -> torch.Tensor:
|
30 |
+
num_experts = b_q_weight.shape[0]
|
31 |
+
assert size_k % 16 == 0
|
32 |
+
output = torch.empty(
|
33 |
+
(num_experts, size_k // 16, size_n * (num_bits // 2)),
|
34 |
+
device=b_q_weight.device,
|
35 |
+
dtype=b_q_weight.dtype,
|
36 |
+
)
|
37 |
+
for e in range(num_experts):
|
38 |
+
output[e] = ops.gptq_marlin_repack(
|
39 |
+
b_q_weight[e], perm[e], size_k, size_n, num_bits
|
40 |
+
)
|
41 |
+
return output
|
42 |
+
|
43 |
+
|
44 |
+
def awq_marlin_moe_repack(
|
45 |
+
b_q_weight: torch.Tensor,
|
46 |
+
perm: torch.Tensor,
|
47 |
+
size_k: int,
|
48 |
+
size_n: int,
|
49 |
+
num_bits: int,
|
50 |
+
) -> torch.Tensor:
|
51 |
+
num_experts = b_q_weight.shape[0]
|
52 |
+
assert size_k % 16 == 0
|
53 |
+
output = torch.empty(
|
54 |
+
(num_experts, size_k // 16, size_n * (num_bits // 2)),
|
55 |
+
device=b_q_weight.device,
|
56 |
+
dtype=b_q_weight.dtype,
|
57 |
+
)
|
58 |
+
for e in range(num_experts):
|
59 |
+
output[e] = ops.awq_marlin_repack(b_q_weight[e], size_k, size_n, num_bits)
|
60 |
+
return output
|
61 |
+
|
62 |
+
|
63 |
+
def moe_sum(input: torch.Tensor, output: torch.Tensor):
|
64 |
+
ops.moe_sum(input, output)
|
65 |
+
|
66 |
+
|
67 |
+
def moe_align_block_size(
|
68 |
+
topk_ids: torch.Tensor,
|
69 |
+
num_experts: int,
|
70 |
+
block_size: int,
|
71 |
+
sorted_token_ids: torch.Tensor,
|
72 |
+
experts_ids: torch.Tensor,
|
73 |
+
num_tokens_post_pad: torch.Tensor,
|
74 |
+
) -> None:
|
75 |
+
ops.moe_align_block_size(
|
76 |
+
topk_ids,
|
77 |
+
num_experts,
|
78 |
+
block_size,
|
79 |
+
sorted_token_ids,
|
80 |
+
experts_ids,
|
81 |
+
num_tokens_post_pad,
|
82 |
+
)
|
83 |
+
|
84 |
+
|
85 |
+
def topk_softmax(
|
86 |
+
topk_weights: torch.Tensor,
|
87 |
+
topk_ids: torch.Tensor,
|
88 |
+
token_expert_indicies: torch.Tensor,
|
89 |
+
gating_output: float,
|
90 |
+
) -> None:
|
91 |
+
ops.topk_softmax(topk_weights, topk_ids, token_expert_indicies, gating_output)
|
92 |
+
|
93 |
+
|
94 |
+
if hasattr(ops, "marlin_gemm_moe"):
|
95 |
+
|
96 |
+
@register_fake(add_op_namespace_prefix("marlin_gemm_moe"))
|
97 |
+
def marlin_gemm_moe_fake(
|
98 |
+
a: torch.Tensor,
|
99 |
+
b_q_weights: torch.Tensor,
|
100 |
+
sorted_ids: torch.Tensor,
|
101 |
+
topk_weights: torch.Tensor,
|
102 |
+
topk_ids: torch.Tensor,
|
103 |
+
b_scales: torch.Tensor,
|
104 |
+
b_zero_points: torch.Tensor,
|
105 |
+
g_idx: torch.Tensor,
|
106 |
+
perm: torch.Tensor,
|
107 |
+
workspace: torch.Tensor,
|
108 |
+
b_q_type: ScalarType,
|
109 |
+
size_m: torch.SymInt,
|
110 |
+
size_n: torch.SymInt,
|
111 |
+
size_k: torch.SymInt,
|
112 |
+
is_k_full: bool,
|
113 |
+
num_experts: int,
|
114 |
+
topk: int,
|
115 |
+
moe_block_size: int,
|
116 |
+
replicate_input: bool,
|
117 |
+
apply_weights: bool,
|
118 |
+
) -> torch.Tensor:
|
119 |
+
return torch.empty((size_m, topk, size_n), dtype=a.dtype, device=a.device)
|
120 |
+
|
121 |
+
|
122 |
+
def silu_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
|
123 |
+
ops.silu_and_mul(out, x)
|
124 |
+
return out
|
125 |
+
|
126 |
+
|
127 |
+
__all__ = [
|
128 |
+
"gptq_marlin_moe_repack",
|
129 |
+
"awq_marlin_moe_repack",
|
130 |
+
"fused_marlin_moe",
|
131 |
+
"moe_sum",
|
132 |
+
"moe_align_block_size",
|
133 |
+
"topk_softmax",
|
134 |
+
"fused_moe",
|
135 |
+
]
|
ext-torch/moe/_custom_ops.py
DELETED
@@ -1,135 +0,0 @@
|
|
1 |
-
from typing import TYPE_CHECKING
|
2 |
-
|
3 |
-
import torch
|
4 |
-
|
5 |
-
# neuron has torch version that doesn't even have impl_abstract
|
6 |
-
if TYPE_CHECKING:
|
7 |
-
|
8 |
-
def register_fake(fn):
|
9 |
-
return lambda name: fn
|
10 |
-
|
11 |
-
else:
|
12 |
-
try:
|
13 |
-
from torch.library import register_fake
|
14 |
-
except ImportError:
|
15 |
-
from torch.library import impl_abstract as register_fake
|
16 |
-
|
17 |
-
try:
|
18 |
-
from ._ops import ops, add_op_namespace_prefix
|
19 |
-
except ImportError as e:
|
20 |
-
# Fallback for local development.
|
21 |
-
try:
|
22 |
-
import _moe
|
23 |
-
|
24 |
-
ops = torch._moe
|
25 |
-
|
26 |
-
def add_op_namespace_prefix(op_name: str):
|
27 |
-
return f"_quantization::{op_name}"
|
28 |
-
|
29 |
-
except ImportError:
|
30 |
-
raise e
|
31 |
-
|
32 |
-
from .scalar_type import ScalarType
|
33 |
-
|
34 |
-
def gptq_marlin_moe_repack(
|
35 |
-
b_q_weight: torch.Tensor,
|
36 |
-
perm: torch.Tensor,
|
37 |
-
size_k: int,
|
38 |
-
size_n: int,
|
39 |
-
num_bits: int,
|
40 |
-
) -> torch.Tensor:
|
41 |
-
num_experts = b_q_weight.shape[0]
|
42 |
-
assert size_k % 16 == 0
|
43 |
-
output = torch.empty(
|
44 |
-
(num_experts, size_k // 16, size_n * (num_bits // 2)),
|
45 |
-
device=b_q_weight.device,
|
46 |
-
dtype=b_q_weight.dtype,
|
47 |
-
)
|
48 |
-
for e in range(num_experts):
|
49 |
-
output[e] = ops.gptq_marlin_repack(
|
50 |
-
b_q_weight[e], perm[e], size_k, size_n, num_bits
|
51 |
-
)
|
52 |
-
return output
|
53 |
-
|
54 |
-
|
55 |
-
def awq_marlin_moe_repack(
|
56 |
-
b_q_weight: torch.Tensor,
|
57 |
-
perm: torch.Tensor,
|
58 |
-
size_k: int,
|
59 |
-
size_n: int,
|
60 |
-
num_bits: int,
|
61 |
-
) -> torch.Tensor:
|
62 |
-
num_experts = b_q_weight.shape[0]
|
63 |
-
assert size_k % 16 == 0
|
64 |
-
output = torch.empty(
|
65 |
-
(num_experts, size_k // 16, size_n * (num_bits // 2)),
|
66 |
-
device=b_q_weight.device,
|
67 |
-
dtype=b_q_weight.dtype,
|
68 |
-
)
|
69 |
-
for e in range(num_experts):
|
70 |
-
output[e] = ops.awq_marlin_repack(b_q_weight[e], size_k, size_n, num_bits)
|
71 |
-
return output
|
72 |
-
|
73 |
-
|
74 |
-
def moe_sum(input: torch.Tensor, output: torch.Tensor):
|
75 |
-
ops.moe_sum(input, output)
|
76 |
-
|
77 |
-
|
78 |
-
def moe_align_block_size(
|
79 |
-
topk_ids: torch.Tensor,
|
80 |
-
num_experts: int,
|
81 |
-
block_size: int,
|
82 |
-
sorted_token_ids: torch.Tensor,
|
83 |
-
experts_ids: torch.Tensor,
|
84 |
-
num_tokens_post_pad: torch.Tensor,
|
85 |
-
) -> None:
|
86 |
-
ops.moe_align_block_size(
|
87 |
-
topk_ids,
|
88 |
-
num_experts,
|
89 |
-
block_size,
|
90 |
-
sorted_token_ids,
|
91 |
-
experts_ids,
|
92 |
-
num_tokens_post_pad,
|
93 |
-
)
|
94 |
-
|
95 |
-
|
96 |
-
def topk_softmax(
|
97 |
-
topk_weights: torch.Tensor,
|
98 |
-
topk_ids: torch.Tensor,
|
99 |
-
token_expert_indicies: torch.Tensor,
|
100 |
-
gating_output: float,
|
101 |
-
) -> None:
|
102 |
-
ops.topk_softmax(topk_weights, topk_ids, token_expert_indicies, gating_output)
|
103 |
-
|
104 |
-
if hasattr(ops, "marlin_gemm_moe"):
|
105 |
-
|
106 |
-
@register_fake(add_op_namespace_prefix("marlin_gemm_moe"))
|
107 |
-
def marlin_gemm_moe_fake(
|
108 |
-
a: torch.Tensor,
|
109 |
-
b_q_weights: torch.Tensor,
|
110 |
-
sorted_ids: torch.Tensor,
|
111 |
-
topk_weights: torch.Tensor,
|
112 |
-
topk_ids: torch.Tensor,
|
113 |
-
b_scales: torch.Tensor,
|
114 |
-
b_zero_points: torch.Tensor,
|
115 |
-
g_idx: torch.Tensor,
|
116 |
-
perm: torch.Tensor,
|
117 |
-
workspace: torch.Tensor,
|
118 |
-
b_q_type: ScalarType,
|
119 |
-
size_m: torch.SymInt,
|
120 |
-
size_n: torch.SymInt,
|
121 |
-
size_k: torch.SymInt,
|
122 |
-
is_k_full: bool,
|
123 |
-
num_experts: int,
|
124 |
-
topk: int,
|
125 |
-
moe_block_size: int,
|
126 |
-
replicate_input: bool,
|
127 |
-
apply_weights: bool,
|
128 |
-
) -> torch.Tensor:
|
129 |
-
return torch.empty((size_m, topk, size_n), dtype=a.dtype, device=a.device)
|
130 |
-
|
131 |
-
|
132 |
-
|
133 |
-
def silu_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
|
134 |
-
ops.silu_and_mul(out, x)
|
135 |
-
return out
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
ext-torch/moe/fused_marlin_moe.py
CHANGED
@@ -7,7 +7,7 @@ import torch
|
|
7 |
|
8 |
from .fused_moe import fused_topk, moe_align_block_size, try_get_optimal_moe_config
|
9 |
from .scalar_type import scalar_types
|
10 |
-
import moe
|
11 |
|
12 |
|
13 |
def get_scalar_type(num_bits: int, has_zp: bool):
|
|
|
7 |
|
8 |
from .fused_moe import fused_topk, moe_align_block_size, try_get_optimal_moe_config
|
9 |
from .scalar_type import scalar_types
|
10 |
+
import moe as ops
|
11 |
|
12 |
|
13 |
def get_scalar_type(num_bits: int, has_zp: bool):
|
ext-torch/moe/fused_moe.py
CHANGED
@@ -11,7 +11,7 @@ import triton.language as tl
|
|
11 |
|
12 |
from .platforms import current_platform
|
13 |
from .fp8 import scaled_fp8_quant
|
14 |
-
import moe
|
15 |
|
16 |
VLLM_FUSED_MOE_CHUNK_SIZE = int(os.getenv("VLLM_FUSED_MOE_CHUNK_SIZE", "32768"))
|
17 |
|
|
|
11 |
|
12 |
from .platforms import current_platform
|
13 |
from .fp8 import scaled_fp8_quant
|
14 |
+
import moe as ops
|
15 |
|
16 |
VLLM_FUSED_MOE_CHUNK_SIZE = int(os.getenv("VLLM_FUSED_MOE_CHUNK_SIZE", "32768"))
|
17 |
|