diff --git a/build.toml b/build.toml
index 90bc8e71aca6e7dfe65d3cd3f9ad5c86ccb52b68..5fe25ad52bc704ca277ddb62960639a69e4b01f2 100644
--- a/build.toml
+++ b/build.toml
@@ -31,6 +31,8 @@ src = [
   "cuda_compat.h",
   "dispatch_utils.h",
   "moe/moe_align_sum_kernels.cu",
+  "moe/moe_wna16.cu",
+  "moe/moe_wna16_utils.h",
   "moe/topk_softmax_kernels.cu",
 ]
 depends = ["torch"]
diff --git a/moe/moe_align_sum_kernels.cu b/moe/moe_align_sum_kernels.cu
index ff74a42d7e810600ab3082902c821b2759f20aea..d7be769458e35b175d557dc62026b4afdf88bc5e 100644
--- a/moe/moe_align_sum_kernels.cu
+++ b/moe/moe_align_sum_kernels.cu
@@ -3,7 +3,7 @@
 #include <c10/cuda/CUDAGuard.h>
 
 #include <ATen/ATen.h>
-#include <THC/THCAtomics.cuh>
+#include <ATen/cuda/Atomic.cuh>
 
 #include "../cuda_compat.h"
 #include "../dispatch_utils.h"
@@ -198,26 +198,27 @@ __global__ void moe_align_block_size_global_mem_kernel(
 }
 
 // taken from
-// https://github.com/sgl-project/sglang/commit/ded9fcd09a43d5e7d5bb31a2bc3e9fc21bf65d2a
+// https://github.com/sgl-project/sglang/commit/cdae77b03dfc6fec3863630550b45bbfc789f957
 template <typename scalar_t>
 __global__ void sgl_moe_align_block_size_kernel(
     scalar_t* __restrict__ topk_ids, int32_t* sorted_token_ids,
     int32_t* expert_ids, int32_t* total_tokens_post_pad, int32_t num_experts,
     int32_t block_size, size_t numel, int32_t* cumsum) {
   __shared__ int32_t shared_counts[32][8];
-  __shared__ int32_t local_offsets[256];
 
-  const int warp_id = threadIdx.x / WARP_SIZE;
-  const int lane_id = threadIdx.x % WARP_SIZE;
+  const int warp_id = threadIdx.x / 32;
   const int experts_per_warp = 8;
   const int my_expert_start = warp_id * experts_per_warp;
 
+  // Initialize shared_counts for this warp's experts
   for (int i = 0; i < experts_per_warp; ++i) {
     if (my_expert_start + i < num_experts) {
       shared_counts[warp_id][i] = 0;
     }
   }
 
+  __syncthreads();
+
   const size_t tokens_per_thread = CEILDIV(numel, blockDim.x);
   const size_t start_idx = threadIdx.x * tokens_per_thread;
 
@@ -230,6 +231,7 @@ __global__ void sgl_moe_align_block_size_kernel(
 
   __syncthreads();
 
+  // Single thread computes cumulative sum and total tokens
   if (threadIdx.x == 0) {
     cumsum[0] = 0;
     for (int i = 1; i <= num_experts; ++i) {
@@ -246,19 +248,28 @@ __global__ void sgl_moe_align_block_size_kernel(
 
   __syncthreads();
 
+  // Assign expert IDs to blocks
   if (threadIdx.x < num_experts) {
     for (int i = cumsum[threadIdx.x]; i < cumsum[threadIdx.x + 1];
          i += block_size) {
       expert_ids[i / block_size] = threadIdx.x;
     }
-    local_offsets[threadIdx.x] = cumsum[threadIdx.x];
   }
+}
 
-  __syncthreads();
-
-  for (int i = start_idx; i < numel && i < start_idx + tokens_per_thread; ++i) {
+// taken from
+// https://github.com/sgl-project/sglang/commit/cdae77b03dfc6fec3863630550b45bbfc789f957
+template <typename scalar_t>
+__global__ void sgl_moe_token_sort_kernel(scalar_t* __restrict__ topk_ids,
+                                          int32_t* sorted_token_ids,
+                                          int32_t* cumsum_buffer,
+                                          size_t numel) {
+  const size_t tid = blockIdx.x * blockDim.x + threadIdx.x;
+  const size_t stride = blockDim.x * gridDim.x;
+
+  for (size_t i = tid; i < numel; i += stride) {
     int32_t expert_id = topk_ids[i];
-    int32_t rank_post_pad = atomicAdd(&local_offsets[expert_id], 1);
+    int32_t rank_post_pad = atomicAdd(&cumsum_buffer[expert_id], 1);
     sorted_token_ids[rank_post_pad] = i;
   }
 }
@@ -377,23 +388,34 @@ void sgl_moe_align_block_size(torch::Tensor topk_ids, int64_t num_experts,
                               torch::Tensor experts_ids,
                               torch::Tensor num_tokens_post_pad) {
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+  TORCH_CHECK(num_experts == 256,
+              "sgl_moe_align_block_size kernel only supports deepseek v3.");
+
   VLLM_DISPATCH_INTEGRAL_TYPES(
       topk_ids.scalar_type(), "sgl_moe_align_block_size_kernel", [&] {
-        // calc needed amount of shared mem for `tokens_cnts` and `cumsum`
-        // tensors
+        // calc needed amount of shared mem for `cumsum` tensors
         auto options_int =
             torch::TensorOptions().dtype(torch::kInt).device(topk_ids.device());
-        // torch::Tensor token_cnts_buffer =
-        //     torch::empty({(num_experts + 1) * num_experts}, options_int);
         torch::Tensor cumsum_buffer =
-            torch::empty({num_experts + 1}, options_int);
+            torch::zeros({num_experts + 1}, options_int);
 
-        auto kernel = vllm::moe::sgl_moe_align_block_size_kernel<scalar_t>;
-        kernel<<<1, 1024, 0, stream>>>(
+        auto align_kernel =
+            vllm::moe::sgl_moe_align_block_size_kernel<scalar_t>;
+        align_kernel<<<1, 1024, 0, stream>>>(
             topk_ids.data_ptr<scalar_t>(), sorted_token_ids.data_ptr<int32_t>(),
             experts_ids.data_ptr<int32_t>(),
             num_tokens_post_pad.data_ptr<int32_t>(), num_experts, block_size,
             topk_ids.numel(), cumsum_buffer.data_ptr<int32_t>());
+
+        const int block_threads = 256;
+        const int num_blocks =
+            (topk_ids.numel() + block_threads - 1) / block_threads;
+        const int max_blocks = 65535;
+        const int actual_blocks = std::min(num_blocks, max_blocks);
+        auto sort_kernel = vllm::moe::sgl_moe_token_sort_kernel<scalar_t>;
+        sort_kernel<<<actual_blocks, block_threads, 0, stream>>>(
+            topk_ids.data_ptr<scalar_t>(), sorted_token_ids.data_ptr<int32_t>(),
+            cumsum_buffer.data_ptr<int32_t>(), topk_ids.numel());
       });
 }
 
diff --git a/moe/moe_wna16.cu b/moe/moe_wna16.cu
new file mode 100644
index 0000000000000000000000000000000000000000..51ae76c1ec882aaac25b2fafabad5abb0a007460
--- /dev/null
+++ b/moe/moe_wna16.cu
@@ -0,0 +1,346 @@
+
+#include <torch/all.h>
+#include <c10/cuda/CUDAGuard.h>
+#include <ATen/cuda/CUDAContext.h>
+#include <cuda_runtime.h>
+
+#include <cuda_fp16.h>
+#include <cuda_bf16.h>
+#include "moe_wna16_utils.h"
+
+#define DIVIDE(x, size) (((x) + (size) - 1) / (size))
+
+template <typename scalar_t, int bit, int GROUPS>
+__global__ void moe_wna16_gemm_kernel(
+    const scalar_t* __restrict__ input, scalar_t* __restrict__ output,
+
+    const uint32_t* __restrict__ qweight, const scalar_t* __restrict__ scales,
+    const uint32_t* __restrict__ qzeros,
+
+    const float* __restrict__ topk_weights,
+    const int32_t* __restrict__ sorted_token_ids,
+    const int32_t* __restrict__ expert_ids,
+    const int32_t* __restrict__ num_tokens_post_pad,
+
+    uint16_t num_experts, uint16_t group_size, uint16_t top_k, uint32_t size_m,
+    uint32_t size_n, uint32_t size_k, uint16_t BLOCK_SIZE_M,
+    uint16_t BLOCK_SIZE_N, uint16_t BLOCK_SIZE_K, bool has_zp,
+    bool mul_topk_weight) {
+#if !defined(__CUDA_ARCH__) || __CUDA_ARCH__ < 800
+  if constexpr (std::is_same<scalar_t, nv_bfloat16>::value) {
+    return;
+  } else {
+#endif
+
+    using Dtype = ScalarType<scalar_t>;
+    using scalar_t2 = typename ScalarType<scalar_t>::scalar_t2;
+
+    if (blockIdx.x * BLOCK_SIZE_M >= num_tokens_post_pad[0]) return;
+
+    const int32_t offset_n = blockIdx.y * BLOCK_SIZE_N + threadIdx.x;
+    const int32_t offset_k = blockIdx.z * BLOCK_SIZE_K;
+
+    const int32_t expert_id = expert_ids[blockIdx.x];
+
+    int32_t num_valid_tokens = 0;
+    extern __shared__ uint16_t block_input_tmp[];
+    scalar_t* block_input = reinterpret_cast<scalar_t*>(block_input_tmp);
+    scalar_t2* block_input_half2 = reinterpret_cast<scalar_t2*>(block_input);
+
+    // load BLOCK_SIZE_M * BLOCK_SIZE_K into shared memory
+    for (int m = 0; m < BLOCK_SIZE_M; m++) {
+      const int32_t offset_m = blockIdx.x * BLOCK_SIZE_M + m;
+      const int32_t token_index = sorted_token_ids[offset_m];
+      if (token_index / top_k >= size_m) break;
+
+      num_valid_tokens = m + 1;
+      if (blockIdx.z == 0 && offset_n < size_n)
+        output[token_index * size_n + offset_n] = Dtype::int2num(0);
+
+      if (expert_id != -1) {
+        int k_per_thread = DIVIDE(BLOCK_SIZE_K, BLOCK_SIZE_N);
+        for (int i = 0; i < k_per_thread; i++) {
+          int k = BLOCK_SIZE_N * i + threadIdx.x;
+          if (k >= BLOCK_SIZE_K) break;
+          if (offset_k + k >= size_k) break;
+
+          // load input to shared memory
+          // use a special layout to fit the layout of dequanted-weight
+          int origin_k;
+          if constexpr (bit == 4) {
+            // [0, 4, 1, 5, 2, 6, 3, 7]
+            int8_t order = (threadIdx.x % 2) * 4 + ((threadIdx.x % 8) / 2);
+            origin_k = BLOCK_SIZE_N * i + threadIdx.x / 8 * 8 + order;
+          } else {
+            // [0, 2, 1, 3]
+            int8_t order = (threadIdx.x % 2) * 2 + ((threadIdx.x % 4) / 2);
+            origin_k = BLOCK_SIZE_N * i + threadIdx.x / 4 * 4 + order;
+          }
+
+          origin_k += token_index / top_k * size_k + blockIdx.z * BLOCK_SIZE_K;
+          block_input[m * BLOCK_SIZE_K + k] = input[origin_k];
+        }
+      }
+    }
+
+    if (expert_id == -1) return;
+    __syncthreads();
+    if (threadIdx.x >= BLOCK_SIZE_N || offset_n >= size_n) return;
+
+    float res[64];  // assume BLOCK_SIZE_M <= 64
+    scalar_t2 res2;
+    scalar_t2 scale_f2;
+    scalar_t2 qzero_f2;
+
+    // note that (size_n * size_k * expert_id) may greater than 2 ** 31
+    constexpr int8_t pack_factor = 32 / bit;
+    const uint64_t expert_offset = ((uint64_t)size_n) * size_k * expert_id;
+    const uint32_t* expert_qweight = qweight + expert_offset / pack_factor;
+    const scalar_t* expert_scales = scales + expert_offset / group_size;
+    const uint32_t* expert_qzeros =
+        qzeros + expert_offset / group_size / pack_factor;
+
+    // load 4*int32 one time: 4 int32 = 128 bit = 1 float4
+    // weight would be loaded in loop
+    uint32_t expert_qweight_tmp[4];
+    float4* expert_qweight_tmp_float4 =
+        reinterpret_cast<float4*>(expert_qweight_tmp);
+
+    // load all required scales one time
+    scalar_t expert_scales_groups[GROUPS];
+    int scales_offset_tmp =
+        (offset_n * size_k + offset_k) / group_size / GROUPS;
+    if constexpr (GROUPS == 1) {
+      *expert_scales_groups = expert_scales[scales_offset_tmp];
+    } else if constexpr (GROUPS == 2) {
+      float* expert_scales_groups_tmp =
+          reinterpret_cast<float*>(expert_scales_groups);
+      *expert_scales_groups_tmp =
+          reinterpret_cast<const float*>(expert_scales)[scales_offset_tmp];
+    } else if constexpr (GROUPS == 4) {
+      float2* expert_scales_groups_tmp =
+          reinterpret_cast<float2*>(expert_scales_groups);
+      *expert_scales_groups_tmp =
+          reinterpret_cast<const float2*>(expert_scales)[scales_offset_tmp];
+    } else if constexpr (GROUPS == 8) {
+      float4* expert_scales_groups_tmp =
+          reinterpret_cast<float4*>(expert_scales_groups);
+      *expert_scales_groups_tmp =
+          reinterpret_cast<const float4*>(expert_scales)[scales_offset_tmp];
+    }
+
+    // load all required qzeros one time
+    uint8_t expert_qzeros_groups[GROUPS];
+    if (!has_zp) {
+      if constexpr (bit == 4) {
+        qzero_f2 = Dtype::num2num2(Dtype::int2num(8));
+      } else {
+        qzero_f2 = Dtype::num2num2(Dtype::int2num(128));
+      }
+    } else {
+      int qzeros_offset_tmp =
+          (offset_n / (8 / bit)) * (size_k / group_size / GROUPS) +
+          offset_k / group_size / GROUPS;
+      if constexpr (GROUPS == 1) {
+        uint8_t* expert_qzeros_groups_tmp =
+            reinterpret_cast<uint8_t*>(expert_qzeros_groups);
+        *expert_qzeros_groups_tmp =
+            reinterpret_cast<const uint8_t*>(expert_qzeros)[qzeros_offset_tmp];
+      } else if constexpr (GROUPS == 2) {
+        uint16_t* expert_qzeros_groups_tmp =
+            reinterpret_cast<uint16_t*>(expert_qzeros_groups);
+        *expert_qzeros_groups_tmp =
+            reinterpret_cast<const uint16_t*>(expert_qzeros)[qzeros_offset_tmp];
+      } else if constexpr (GROUPS == 4) {
+        uint32_t* expert_qzeros_groups_tmp =
+            reinterpret_cast<uint32_t*>(expert_qzeros_groups);
+        *expert_qzeros_groups_tmp =
+            reinterpret_cast<const uint32_t*>(expert_qzeros)[qzeros_offset_tmp];
+      } else if constexpr (GROUPS == 8) {
+        uint64_t* expert_qzeros_groups_tmp =
+            reinterpret_cast<uint64_t*>(expert_qzeros_groups);
+        *expert_qzeros_groups_tmp =
+            reinterpret_cast<const uint64_t*>(expert_qzeros)[qzeros_offset_tmp];
+      }
+    }
+
+    for (int tmp_k = 0; tmp_k < BLOCK_SIZE_K / pack_factor; tmp_k++) {
+      int k = offset_k + tmp_k * pack_factor;
+      if (k >= size_k) break;
+      const int32_t weight_offset = offset_n * size_k + k;
+
+      if (tmp_k % 4 == 0) {
+        *expert_qweight_tmp_float4 = reinterpret_cast<const float4*>(
+            expert_qweight)[weight_offset / pack_factor / 4];
+      }
+
+      if (tmp_k % (group_size / pack_factor) == 0) {
+        scalar_t scale_f =
+            expert_scales_groups[tmp_k / (group_size / pack_factor)];
+        scale_f2 = Dtype::num2num2(scale_f);
+
+        if (has_zp) {
+          uint8_t qzero =
+              expert_qzeros_groups[tmp_k / (group_size / pack_factor)];
+          if constexpr (bit == 4) {
+            qzero = (qzero >> ((threadIdx.x % 2) * 4)) & 0xF;
+          }
+          qzero_f2 = Dtype::num2num2(Dtype::int2num(qzero));
+        }
+      }
+
+      scalar_t2 weight_half2[16 / bit];
+      dequant<scalar_t2, bit>(expert_qweight_tmp[tmp_k % 4], weight_half2);
+
+      for (int m = 0; m < num_valid_tokens; m++) {
+        res2 = {};
+
+#pragma unroll
+        for (int i = 0; i < 16 / bit; i++) {
+          int32_t offset_input = m * BLOCK_SIZE_K / 2 + tmp_k * (16 / bit) + i;
+          res2 = __hfma2(__hmul2(__hsub2(weight_half2[i], qzero_f2), scale_f2),
+                         block_input_half2[offset_input], res2);
+        }
+
+        if (tmp_k == 0) {
+          res[m] = Dtype::num2float(res2.x) + Dtype::num2float(res2.y);
+        } else {
+          res[m] += Dtype::num2float(res2.x) + Dtype::num2float(res2.y);
+        }
+      }
+    }
+
+    for (int m = 0; m < num_valid_tokens; ++m) {
+      const int32_t token_index =
+          sorted_token_ids[blockIdx.x * BLOCK_SIZE_M + m];
+      if (mul_topk_weight) {
+        res[m] *= topk_weights[token_index];
+      }
+      atomicAdd(&output[token_index * size_n + offset_n],
+                Dtype::float2num(res[m]));
+    }
+
+#if !defined(__CUDA_ARCH__) || __CUDA_ARCH__ < 800
+  }
+#endif
+}
+
+template <typename scalar_t>
+void run_moe_wna16_gemm(const scalar_t* input, scalar_t* output,
+                        const uint32_t* b_qweight, const scalar_t* b_scales,
+                        const uint32_t* b_qzeros, const float* topk_weights,
+                        const int32_t* sorted_token_ids,
+                        const int32_t* expert_ids,
+                        const int32_t* num_tokens_post_pad, int num_experts,
+                        int group_size, int num_token_blocks, int top_k,
+                        int size_m, int size_n, int size_k, int BLOCK_SIZE_M,
+                        int BLOCK_SIZE_N, int BLOCK_SIZE_K, int bit,
+                        bool has_zp, bool mul_topk_weight) {
+  dim3 blockDim, gridDim;
+  blockDim.x = BLOCK_SIZE_N;
+  blockDim.y = 1;
+  blockDim.z = 1;
+  gridDim.x = num_token_blocks;
+  gridDim.y = DIVIDE(size_n, BLOCK_SIZE_N);
+  gridDim.z = DIVIDE(size_k, BLOCK_SIZE_K);
+
+  auto kernel = moe_wna16_gemm_kernel<scalar_t, 4, 1>;
+  if (bit == 4) {
+    if (BLOCK_SIZE_K / group_size == 2) {
+      kernel = moe_wna16_gemm_kernel<scalar_t, 4, 2>;
+    } else if (BLOCK_SIZE_K / group_size == 4) {
+      kernel = moe_wna16_gemm_kernel<scalar_t, 4, 4>;
+    } else if (BLOCK_SIZE_K / group_size == 8) {
+      kernel = moe_wna16_gemm_kernel<scalar_t, 4, 8>;
+    }
+  } else {
+    if (BLOCK_SIZE_K / group_size == 1) {
+      kernel = moe_wna16_gemm_kernel<scalar_t, 8, 1>;
+    } else if (BLOCK_SIZE_K / group_size == 2) {
+      kernel = moe_wna16_gemm_kernel<scalar_t, 8, 2>;
+    } else if (BLOCK_SIZE_K / group_size == 4) {
+      kernel = moe_wna16_gemm_kernel<scalar_t, 8, 4>;
+    } else if (BLOCK_SIZE_K / group_size == 8) {
+      kernel = moe_wna16_gemm_kernel<scalar_t, 8, 8>;
+    }
+  }
+
+  const int shared_mem_size = BLOCK_SIZE_M * BLOCK_SIZE_K * 2;
+  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+  kernel<<<gridDim, blockDim, shared_mem_size, stream>>>(
+      input, output, b_qweight, b_scales, b_qzeros, topk_weights,
+      sorted_token_ids, expert_ids, num_tokens_post_pad, num_experts,
+      group_size, top_k, size_m, size_n, size_k, BLOCK_SIZE_M, BLOCK_SIZE_N,
+      BLOCK_SIZE_K, has_zp, mul_topk_weight);
+}
+
+torch::Tensor moe_wna16_gemm(torch::Tensor input, torch::Tensor output,
+                             torch::Tensor b_qweight, torch::Tensor b_scales,
+                             std::optional<torch::Tensor> b_qzeros,
+                             std::optional<torch::Tensor> topk_weights,
+                             torch::Tensor sorted_token_ids,
+                             torch::Tensor expert_ids,
+                             torch::Tensor num_tokens_post_pad, int64_t top_k,
+                             int64_t BLOCK_SIZE_M, int64_t BLOCK_SIZE_N,
+                             int64_t BLOCK_SIZE_K, int64_t bit) {
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(input));
+  auto options =
+      torch::TensorOptions().dtype(input.dtype()).device(input.device());
+
+  const int num_experts = b_qweight.size(0);
+  const int size_m = input.size(0);
+  const int size_n = b_qweight.size(1);
+  const int size_k = input.size(1);
+  const int group_size = size_k / b_scales.size(2);
+
+  int64_t EM = sorted_token_ids.size(0);
+  if (size_m <= BLOCK_SIZE_M) {
+    EM = min(EM, size_m * BLOCK_SIZE_M * top_k);
+  }
+  const int num_token_blocks = (EM + BLOCK_SIZE_M - 1) / BLOCK_SIZE_M;
+
+  const uint32_t* b_qzeros_ptr;
+  if (b_qzeros.has_value())
+    b_qzeros_ptr = (const uint32_t*)b_qzeros.value().data_ptr<uint8_t>();
+  const float* topk_weights_ptr;
+  if (topk_weights.has_value())
+    topk_weights_ptr = (const float*)topk_weights.value().data_ptr();
+
+  int groups_per_block_row = BLOCK_SIZE_K / group_size;
+  TORCH_CHECK(bit == 4 || bit == 8, "bit must be 4 or 8");
+  TORCH_CHECK(size_k % BLOCK_SIZE_K == 0,
+              "size_k must divisible by BLOCK_SIZE_K");
+  TORCH_CHECK(BLOCK_SIZE_K % group_size == 0,
+              "BLOCK_SIZE_K must divisible by group_size");
+  TORCH_CHECK(BLOCK_SIZE_M <= 64, "BLOCK_SIZE_M must less or equal to 64");
+  TORCH_CHECK(groups_per_block_row == 1 || groups_per_block_row == 2 ||
+                  groups_per_block_row == 4 || groups_per_block_row == 8,
+              "BLOCK_SIZE_K // group_size must be one of [1, 2, 4, 8]");
+
+  if (input.scalar_type() == at::ScalarType::Half) {
+    run_moe_wna16_gemm<half>(
+        (const half*)input.data_ptr<at::Half>(),
+        (half*)output.data_ptr<at::Half>(),
+        (const uint32_t*)b_qweight.data_ptr<uint8_t>(),
+        (const half*)b_scales.data_ptr<at::Half>(), b_qzeros_ptr,
+        topk_weights_ptr, sorted_token_ids.data_ptr<int32_t>(),
+        expert_ids.data_ptr<int32_t>(), num_tokens_post_pad.data_ptr<int32_t>(),
+        num_experts, group_size, num_token_blocks, top_k, size_m, size_n,
+        size_k, BLOCK_SIZE_M, BLOCK_SIZE_N, BLOCK_SIZE_K, bit,
+        b_qzeros.has_value(), topk_weights.has_value());
+  } else if (input.scalar_type() == at::ScalarType::BFloat16) {
+    run_moe_wna16_gemm<nv_bfloat16>(
+        (const nv_bfloat16*)input.data_ptr<at::BFloat16>(),
+        (nv_bfloat16*)output.data_ptr<at::BFloat16>(),
+        (const uint32_t*)b_qweight.data_ptr<uint8_t>(),
+        (const nv_bfloat16*)b_scales.data_ptr<at::BFloat16>(), b_qzeros_ptr,
+        topk_weights_ptr, sorted_token_ids.data_ptr<int32_t>(),
+        expert_ids.data_ptr<int32_t>(), num_tokens_post_pad.data_ptr<int32_t>(),
+        num_experts, group_size, num_token_blocks, top_k, size_m, size_n,
+        size_k, BLOCK_SIZE_M, BLOCK_SIZE_N, BLOCK_SIZE_K, bit,
+        b_qzeros.has_value(), topk_weights.has_value());
+  } else {
+    TORCH_CHECK(false, "moe_wna16_gemm only supports bfloat16 and float16");
+  }
+  return output;
+}
diff --git a/moe/moe_wna16_utils.h b/moe/moe_wna16_utils.h
new file mode 100644
index 0000000000000000000000000000000000000000..4396b80240efe98741eeb85586ea2b1f9146a17b
--- /dev/null
+++ b/moe/moe_wna16_utils.h
@@ -0,0 +1,200 @@
+
+#include <cuda_fp16.h>
+#include <cuda_bf16.h>
+
+template <typename scalar_t>
+class ScalarType {};
+
+template <>
+class ScalarType<half> {
+ public:
+  using scalar_t = half;
+  using scalar_t2 = half2;
+
+  static __device__ float inline num2float(const half x) {
+    return __half2float(x);
+  }
+
+  static __device__ half2 inline num2num2(const half x) {
+    return __half2half2(x);
+  }
+
+  static __device__ half2 inline nums2num2(const half x1, const half x2) {
+    return __halves2half2(x1, x2);
+  }
+
+  static __host__ __device__ half inline float2num(const float x) {
+    return __float2half(x);
+  }
+
+  static __host__ __device__ half inline int2num(const float x) {
+    return __int2half_rn(x);
+  }
+
+  static __host__ __device__ float2 inline num22float2(const half2 x) {
+    return __half22float2(x);
+  }
+
+  static __host__ __device__ half2 inline float22num2(const float2 x) {
+    return __float22half2_rn(x);
+  }
+};
+
+template <>
+class ScalarType<nv_bfloat16> {
+ public:
+  using scalar_t = nv_bfloat16;
+  using scalar_t2 = nv_bfloat162;
+
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800
+  static __device__ float inline num2float(const nv_bfloat16 x) {
+    return __bfloat162float(x);
+  }
+
+  static __device__ nv_bfloat162 inline num2num2(const nv_bfloat16 x) {
+    return __bfloat162bfloat162(x);
+  }
+
+  static __device__ nv_bfloat162 inline nums2num2(const nv_bfloat16 x1,
+                                                  const nv_bfloat16 x2) {
+    return __halves2bfloat162(x1, x2);
+  }
+
+  static __host__ __device__ nv_bfloat16 inline float2num(const float x) {
+    return __float2bfloat16(x);
+  }
+
+  static __host__ __device__ nv_bfloat16 inline int2num(const float x) {
+    return __int2bfloat16_rn(x);
+  }
+
+  static __host__ __device__ float2 inline num22float2(const nv_bfloat162 x) {
+    return __bfloat1622float2(x);
+  }
+
+  static __host__ __device__ nv_bfloat162 inline float22num2(const float2 x) {
+    return __float22bfloat162_rn(x);
+  }
+#endif
+};
+
+template <int lut>
+__device__ inline int lop3(int a, int b, int c) {
+  int res;
+  asm volatile("lop3.b32 %0, %1, %2, %3, %4;\n"
+               : "=r"(res)
+               : "r"(a), "r"(b), "r"(c), "n"(lut));
+  return res;
+}
+
+template <int start_byte, int mask>
+__device__ inline uint32_t prmt(uint32_t a) {
+  uint32_t res;
+  asm volatile("prmt.b32 %0, %1, %2, %3;\n"
+               : "=r"(res)
+               : "r"(a), "n"(start_byte), "n"(mask));
+  return res;
+}
+
+template <typename scalar_t2, int bit>
+__device__ inline void dequant(int q, scalar_t2* res) {}
+
+template <>
+__device__ inline void dequant<half2, 4>(int q, half2* res) {
+  const int LO = 0x000f000f;
+  const int HI = 0x00f000f0;
+  const int EX = 0x64006400;
+  const int SUB = 0x64006400;
+  const int MUL = 0x2c002c00;
+  const int ADD = 0xd400d400;
+
+  int lo0 = lop3 < (0xf0 & 0xcc) | 0xaa > (q, LO, EX);
+  int hi0 = lop3 < (0xf0 & 0xcc) | 0xaa > (q, HI, EX);
+  q >>= 8;
+  int lo1 = lop3 < (0xf0 & 0xcc) | 0xaa > (q, LO, EX);
+  int hi1 = lop3 < (0xf0 & 0xcc) | 0xaa > (q, HI, EX);
+
+  res[0] = __hsub2(*reinterpret_cast<half2*>(&lo0),
+                   *reinterpret_cast<const half2*>(&SUB));
+  res[1] = __hfma2(*reinterpret_cast<half2*>(&hi0),
+                   *reinterpret_cast<const half2*>(&MUL),
+                   *reinterpret_cast<const half2*>(&ADD));
+  res[2] = __hsub2(*reinterpret_cast<half2*>(&lo1),
+                   *reinterpret_cast<const half2*>(&SUB));
+  res[3] = __hfma2(*reinterpret_cast<half2*>(&hi1),
+                   *reinterpret_cast<const half2*>(&MUL),
+                   *reinterpret_cast<const half2*>(&ADD));
+}
+
+template <>
+__device__ inline void dequant<half2, 8>(int q, half2* res) {
+  static constexpr uint32_t mask_for_elt_01 = 0x5250;
+  static constexpr uint32_t mask_for_elt_23 = 0x5351;
+  static constexpr uint32_t start_byte_for_fp16 = 0x64646464;
+
+  uint32_t lo = prmt<start_byte_for_fp16, mask_for_elt_01>(q);
+  uint32_t hi = prmt<start_byte_for_fp16, mask_for_elt_23>(q);
+
+  static constexpr uint32_t I8s_TO_F16s_MAGIC_NUM = 0x64006400;
+
+  res[0] = __hsub2(*reinterpret_cast<half2*>(&lo),
+                   *reinterpret_cast<const half2*>(&I8s_TO_F16s_MAGIC_NUM));
+  res[1] = __hsub2(*reinterpret_cast<half2*>(&hi),
+                   *reinterpret_cast<const half2*>(&I8s_TO_F16s_MAGIC_NUM));
+}
+
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800
+template <>
+__device__ inline void dequant<nv_bfloat162, 4>(int q, nv_bfloat162* res) {
+  static constexpr uint32_t MASK = 0x000f000f;
+  static constexpr uint32_t EX = 0x43004300;
+
+  int lo0 = lop3 < (0xf0 & 0xcc) | 0xaa > (q, MASK, EX);
+  q >>= 4;
+  int hi0 = lop3 < (0xf0 & 0xcc) | 0xaa > (q, MASK, EX);
+  q >>= 4;
+  int lo1 = lop3 < (0xf0 & 0xcc) | 0xaa > (q, MASK, EX);
+  q >>= 4;
+  int hi1 = lop3 < (0xf0 & 0xcc) | 0xaa > (q, MASK, EX);
+
+  static constexpr uint32_t MUL = 0x3F803F80;
+  static constexpr uint32_t ADD = 0xC300C300;
+
+  res[0] = __hfma2(*reinterpret_cast<nv_bfloat162*>(&lo0),
+                   *reinterpret_cast<const nv_bfloat162*>(&MUL),
+                   *reinterpret_cast<const nv_bfloat162*>(&ADD));
+  res[1] = __hfma2(*reinterpret_cast<nv_bfloat162*>(&hi0),
+                   *reinterpret_cast<const nv_bfloat162*>(&MUL),
+                   *reinterpret_cast<const nv_bfloat162*>(&ADD));
+  res[2] = __hfma2(*reinterpret_cast<nv_bfloat162*>(&lo1),
+                   *reinterpret_cast<const nv_bfloat162*>(&MUL),
+                   *reinterpret_cast<const nv_bfloat162*>(&ADD));
+  res[3] = __hfma2(*reinterpret_cast<nv_bfloat162*>(&hi1),
+                   *reinterpret_cast<const nv_bfloat162*>(&MUL),
+                   *reinterpret_cast<const nv_bfloat162*>(&ADD));
+}
+
+template <>
+__device__ inline void dequant<nv_bfloat162, 8>(int q, nv_bfloat162* res) {
+  float fp32_intermediates[4];
+  uint32_t* fp32_intermediates_casted =
+      reinterpret_cast<uint32_t*>(fp32_intermediates);
+
+  static constexpr uint32_t fp32_base = 0x4B000000;
+  fp32_intermediates_casted[0] = __byte_perm(q, fp32_base, 0x7650);
+  fp32_intermediates_casted[1] = __byte_perm(q, fp32_base, 0x7652);
+  fp32_intermediates_casted[2] = __byte_perm(q, fp32_base, 0x7651);
+  fp32_intermediates_casted[3] = __byte_perm(q, fp32_base, 0x7653);
+
+  fp32_intermediates[0] -= 8388608.f;
+  fp32_intermediates[1] -= 8388608.f;
+  fp32_intermediates[2] -= 8388608.f;
+  fp32_intermediates[3] -= 8388608.f;
+
+  uint32_t* bf16_result_ptr = reinterpret_cast<uint32_t*>(res);
+  bf16_result_ptr[0] = __byte_perm(fp32_intermediates_casted[0],
+                                   fp32_intermediates_casted[1], 0x7632);
+  bf16_result_ptr[1] = __byte_perm(fp32_intermediates_casted[2],
+                                   fp32_intermediates_casted[3], 0x7632);
+}
+#endif
diff --git a/tests/kernels/test_moe.py b/tests/kernels/test_moe.py
index 0994b499e06bda97fed1a2bad4e62f80216058c7..9c3ae0c2efb6c68dc9a235e260e790318e1264c7 100644
--- a/tests/kernels/test_moe.py
+++ b/tests/kernels/test_moe.py
@@ -1,3 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
 """Tests for the MOE layers.
 
 Run `pytest tests/kernels/test_moe.py`.
@@ -18,12 +19,15 @@ from moe.utils.marlin_utils_test import marlin_quantize, quantize_weights
 from .utils import compute_max_diff, opcheck, torch_moe
 
 
+from torch.nn import Parameter
+from torch.nn import functional as F
+
 def stack_and_dev(tensors: List[torch.Tensor]):
     dev = tensors[0].device
     return torch.stack(tensors, dim=0).to(dev)
 
-
 NUM_EXPERTS = [8, 64]
+EP_SIZE = [1, 4]
 TOP_KS = [2, 6]
 
 
@@ -32,25 +36,54 @@ TOP_KS = [2, 6]
 @pytest.mark.parametrize("k", [128, 511, 1024])
 @pytest.mark.parametrize("e", NUM_EXPERTS)
 @pytest.mark.parametrize("topk", TOP_KS)
+@pytest.mark.parametrize("ep_size", EP_SIZE)
 @pytest.mark.parametrize("dtype", [torch.float16, torch.bfloat16])
+@pytest.mark.parametrize("padding", [True, False])
 def test_fused_moe(
     m: int,
     n: int,
     k: int,
     e: int,
     topk: int,
+    ep_size: int,
     dtype: torch.dtype,
+    padding: bool,
 ):
     a = torch.randn((m, k), device="cuda", dtype=dtype) / 10
     w1 = torch.randn((e, 2 * n, k), device="cuda", dtype=dtype) / 10
     w2 = torch.randn((e, k, n), device="cuda", dtype=dtype) / 10
 
     score = torch.randn((m, e), device="cuda", dtype=dtype)
-    triton_output = fused_moe(a, w1, w2, score, topk, renormalize=False)
-    torch_output = torch_moe(a, w1, w2, score, topk)
+
+    if ep_size > 1:
+        local_e = e // ep_size
+        e_ids = torch.randint(0,
+                              e, (local_e, ),
+                              device="cuda",
+                              dtype=torch.int32)
+        e_map = torch.full((e, ), -1, device="cuda", dtype=torch.int32)
+        e_map[e_ids] = torch.arange(local_e, device="cuda", dtype=torch.int32)
+        w1 = w1[e_ids]
+        w2 = w2[e_ids]
+    else:
+        e_map = None
+
+    torch_output = torch_moe(a, w1, w2, score, topk, e_map)
+    if padding:
+        w1 = F.pad(w1, (0, 128), "constant", 0)[..., 0:-128]
+        torch.cuda.empty_cache()
+        w2 = F.pad(w2, (0, 128), "constant", 0)[..., 0:-128]
+        torch.cuda.empty_cache()
+
+    triton_output = fused_moe(a,
+                              w1,
+                              w2,
+                              score,
+                              topk,
+                              global_num_experts=e,
+                              expert_map=e_map,
+                              renormalize=False)
     torch.testing.assert_close(triton_output, torch_output, atol=2e-2, rtol=0)
-    # iterative_output = iterative_moe(a, w1, w2, score, topk, renormalize=False)
-    # torch.testing.assert_close(iterative_output, torch_output, atol=2e-2, rtol=0)
 
 
 @pytest.mark.parametrize("m", [1, 32, 222])
@@ -58,21 +91,14 @@ def test_fused_moe(
 @pytest.mark.parametrize("k", [128, 1024])
 @pytest.mark.parametrize("e", NUM_EXPERTS)
 @pytest.mark.parametrize("topk", TOP_KS)
+@pytest.mark.parametrize("ep_size", EP_SIZE)
 @pytest.mark.parametrize("dtype", [torch.float16, torch.bfloat16])
 @pytest.mark.parametrize("group_size", [64, 128])
 @pytest.mark.parametrize("has_zp", [True, False])
 @pytest.mark.parametrize("weight_bits", [4, 8])
-def test_fused_moe_wn16(
-    m: int,
-    n: int,
-    k: int,
-    e: int,
-    topk: int,
-    dtype: torch.dtype,
-    group_size: int,
-    has_zp: bool,
-    weight_bits: int,
-):
+def test_fused_moe_wn16(m: int, n: int, k: int, e: int, topk: int,
+                        ep_size: int, dtype: torch.dtype, group_size: int,
+                        has_zp: bool, weight_bits: int):
     print(m, n, k, e, topk, dtype, group_size, has_zp, weight_bits)
     a = torch.randn((m, k), device="cuda", dtype=dtype) / 10
     w1 = torch.randn((e, 2 * n, k), device="cuda", dtype=dtype) / 10
@@ -88,40 +114,35 @@ def test_fused_moe_wn16(
 
     w1_ref = w1.clone()
     w2_ref = w2.clone()
-    w1_qweight = torch.empty(
-        (e, 2 * n, k // pack_factor), device="cuda", dtype=torch.uint8
-    )
-    w2_qweight = torch.empty((e, k, n // pack_factor), device="cuda", dtype=torch.uint8)
-    w1_scales = torch.empty((e, 2 * n, k // group_size), device="cuda", dtype=dtype)
-    w2_scales = torch.empty((e, k, n // group_size), device="cuda", dtype=dtype)
-    w1_qzeros = torch.empty(
-        (e, 2 * n // pack_factor, k // group_size), device="cuda", dtype=torch.uint8
-    )
-    w2_qzeros = torch.empty(
-        (e, k // pack_factor, n // group_size), device="cuda", dtype=torch.uint8
-    )
+    w1_qweight = torch.empty((e, 2 * n, k // pack_factor),
+                             device="cuda",
+                             dtype=torch.uint8)
+    w2_qweight = torch.empty((e, k, n // pack_factor),
+                             device="cuda",
+                             dtype=torch.uint8)
+    w1_scales = torch.empty((e, 2 * n, k // group_size),
+                            device="cuda",
+                            dtype=dtype)
+    w2_scales = torch.empty((e, k, n // group_size),
+                            device="cuda",
+                            dtype=dtype)
+    w1_qzeros = torch.empty((e, 2 * n // pack_factor, k // group_size),
+                            device="cuda",
+                            dtype=torch.uint8)
+    w2_qzeros = torch.empty((e, k // pack_factor, n // group_size),
+                            device="cuda",
+                            dtype=torch.uint8)
 
     for i in range(e * 2):
         expert_id = i % e
         if i // e == 0:
-            w, w_ref, w_qweight, w_scales, w_qzeros = (
-                w1,
-                w1_ref,
-                w1_qweight,
-                w1_scales,
-                w1_qzeros,
-            )
+            w, w_ref, w_qweight, w_scales, w_qzeros = \
+                w1, w1_ref, w1_qweight, w1_scales, w1_qzeros
         else:
-            w, w_ref, w_qweight, w_scales, w_qzeros = (
-                w2,
-                w2_ref,
-                w2_qweight,
-                w2_scales,
-                w2_qzeros,
-            )
+            w, w_ref, w_qweight, w_scales, w_qzeros = \
+                w2, w2_ref, w2_qweight, w2_scales, w2_qzeros
         weight, qweight, scales, qzeros = quantize_weights(
-            w[expert_id].T, quant_type, group_size, has_zp, False
-        )
+            w[expert_id].T, quant_type, group_size, has_zp, False)
         weight = weight.T
         qweight = qweight.T.contiguous().to(torch.uint8)
         scales = scales.T
@@ -138,25 +159,45 @@ def test_fused_moe_wn16(
         if has_zp:
             w_qzeros[expert_id] = qzeros
 
-    triton_output = fused_moe(
-        a,
-        w1_qweight,
-        w2_qweight,
-        score,
-        topk,
-        renormalize=False,
-        use_int4_w4a16=weight_bits == 4,
-        use_int8_w8a16=weight_bits == 8,
-        w1_scale=w1_scales,
-        w2_scale=w2_scales,
-        w1_zp=w1_qzeros if has_zp else None,
-        w2_zp=w2_qzeros if has_zp else None,
-        block_shape=[0, group_size],
-    )
-    torch_output = torch_moe(a, w1_ref, w2_ref, score, topk)
+    if ep_size > 1:
+        local_e = e // ep_size
+        e_ids = torch.randint(0,
+                              e, (local_e, ),
+                              device="cuda",
+                              dtype=torch.int32)
+        e_map = torch.full((e, ), -1, device="cuda", dtype=torch.int32)
+        e_map[e_ids] = torch.arange(local_e, device="cuda", dtype=torch.int32)
+        w1_ref = w1_ref[e_ids]
+        w2_ref = w2_ref[e_ids]
+        w1_qweight = w1_qweight[e_ids]
+        w2_qweight = w2_qweight[e_ids]
+        w1_scales = w1_scales[e_ids]
+        w2_scales = w2_scales[e_ids]
+        w1_qzeros = w1_qzeros[e_ids]
+        w2_qzeros = w2_qzeros[e_ids]
+    else:
+        e_map = None
+
+    triton_output = fused_moe(a,
+                              w1_qweight,
+                              w2_qweight,
+                              score,
+                              topk,
+                              renormalize=False,
+                              use_int4_w4a16=weight_bits == 4,
+                              use_int8_w8a16=weight_bits == 8,
+                              global_num_experts=e,
+                              expert_map=e_map,
+                              w1_scale=w1_scales,
+                              w2_scale=w2_scales,
+                              w1_zp=w1_qzeros if has_zp else None,
+                              w2_zp=w2_qzeros if has_zp else None,
+                              block_shape=[0, group_size])
+    torch_output = torch_moe(a, w1_ref, w2_ref, score, topk, e_map)
     torch.testing.assert_close(triton_output, torch_output, atol=2e-2, rtol=0)
 
 
+
 @pytest.mark.parametrize("m", [1, 33, 64, 222])
 @pytest.mark.parametrize("n", [128, 2048])
 @pytest.mark.parametrize("k", [128, 1024])
@@ -178,7 +219,7 @@ def test_fused_marlin_moe(
     num_bits: int,
     is_k_full: bool,
 ):
-    torch.manual_seed(7)
+    current_platform.seed_everything(7)
 
     # Filter act_order
     if act_order:
@@ -190,7 +231,8 @@ def test_fused_marlin_moe(
         if not is_k_full:
             return
 
-    quant_type = scalar_types.uint4b8 if num_bits == 4 else scalar_types.uint8b128
+    quant_type = (scalar_types.uint4b8
+                  if num_bits == 4 else scalar_types.uint8b128)
     dtype = torch.float16
     a = torch.randn((m, k), device="cuda", dtype=dtype) / 10
     w1 = torch.randn((e, 2 * n, k), device="cuda", dtype=dtype) / 10
@@ -205,8 +247,8 @@ def test_fused_marlin_moe(
     for i in range(w1.shape[0]):
         test_perm = torch.randperm(k)
         w_ref1, qweight1, scales1, g_idx1, sort_indices1, _ = marlin_quantize(
-            w1[i].transpose(1, 0), quant_type, group_size, act_order, test_perm
-        )
+            w1[i].transpose(1, 0), quant_type, group_size, act_order,
+            test_perm)
         w_ref1_l.append(w_ref1)
         qweight1_l.append(qweight1)
         scales1_l.append(scales1)
@@ -228,8 +270,8 @@ def test_fused_marlin_moe(
     for i in range(w2.shape[0]):
         test_perm = torch.randperm(n)
         w_ref2, qweight2, scales2, g_idx2, sort_indices2, _ = marlin_quantize(
-            w2[i].transpose(1, 0), quant_type, group_size, act_order, test_perm
-        )
+            w2[i].transpose(1, 0), quant_type, group_size, act_order,
+            test_perm)
         w_ref2_l.append(w_ref2)
         qweight2_l.append(qweight2)
         scales2_l.append(scales2)
@@ -273,79 +315,141 @@ def test_fused_marlin_moe(
 
     assert compute_max_diff(marlin_output, triton_output) < 4e-2
 
-    token_expert_indicies = torch.empty(m, topk, dtype=torch.int32, device=a.device)
+    token_expert_indicies = torch.empty(m,
+                                        topk,
+                                        dtype=torch.int32,
+                                        device=a.device)
 
-    opcheck(
-        ops.topk_softmax,
-        (
-            topk_weights,
-            topk_ids,
-            token_expert_indicies,
-            score.float(),
-        ),
-    )
+    opcheck(ops.topk_softmax, (
+        topk_weights,
+        topk_ids,
+        token_expert_indicies,
+        score.float(),
+    ))
 
     block_size_m = 4
 
-    sorted_token_ids, _, _ = moe_align_block_size(topk_ids, block_size_m, e)
+    sorted_token_ids, _, _ = moe_align_block_size(topk_ids, block_size_m,
+                                                    e)
 
     max_workspace_size = ((m + 255) // 256) * (max(2 * n, k) // 64) * 16
-    workspace = torch.zeros(
-        max_workspace_size, dtype=torch.int, device="cuda", requires_grad=False
-    )
+    workspace = torch.zeros(max_workspace_size,
+                            dtype=torch.int,
+                            device="cuda",
+                            requires_grad=False)
+
+    zp = torch.empty((0, 0),
+                        dtype=dtype,
+                        device="cuda",
+                        requires_grad=False)
+    opcheck(ops.marlin_gemm_moe,
+            (a, qweight1, sorted_token_ids, topk_weights, topk_ids,
+                scales1, zp, g_idx1, sort_indices1, workspace, quant_type.id,
+                m, 2 * n, k, True, e, topk, block_size_m, True, False))
+
+
+@pytest.mark.skip("This test is here for the sake of debugging, "
+                  "don't run it in automated tests.")
+@pytest.mark.parametrize("m", [64, 512, 222, 33, 1])
+@pytest.mark.parametrize("n", [128, 2048, 256, 1024])
+@pytest.mark.parametrize("k", [128, 1024, 512])
+@pytest.mark.parametrize("e", [8, 64])
+@pytest.mark.parametrize("topk", [2, 6])
+@pytest.mark.parametrize("group_size", [-1, 32, 64, 128])
+@pytest.mark.parametrize("act_order", [True, False])
+@pytest.mark.parametrize("num_bits", [4, 8])
+@pytest.mark.parametrize("is_k_full", [True, False])
+@pytest.mark.skipif(current_platform.is_rocm(), reason="Skip for rocm")
+def test_single_marlin_moe_multiply(
+    m: int,
+    n: int,
+    k: int,
+    e: int,
+    topk: int,
+    group_size: int,
+    act_order: bool,
+    num_bits: int,
+    is_k_full: bool,
+):
+
+    # Filter act_order
+    if act_order:
+        if group_size == -1:
+            return
+        if group_size == k:
+            return
+    else:
+        if not is_k_full:
+            return
+
+    quant_type = (scalar_types.uint4b8
+                  if num_bits == 4 else scalar_types.uint8b128)
+    dtype = torch.float16
+    a = torch.randn((m, k), device="cuda", dtype=dtype) / 10
+    w = torch.randn((e, n, k), device="cuda", dtype=dtype) / 10
+
+    w_ref_l = []
+    qweights_l = []
+    scales_l = []
+    g_idx_l = []
+    sort_indices_l = []
+
+    for i in range(w.shape[0]):
+        test_perm = torch.randperm(k)
+        w_ref, qweight, scales, g_idx, sort_indices, _ = marlin_quantize(
+            w[i].transpose(1, 0), quant_type, group_size, act_order, test_perm)
+        w_ref_l.append(w_ref)
+        qweights_l.append(qweight)
+        scales_l.append(scales)
+        g_idx_l.append(g_idx)
+        sort_indices_l.append(sort_indices)
+
+    w_ref = stack_and_dev(w_ref_l)
+    qweight = stack_and_dev(qweights_l).contiguous()
+    scales = stack_and_dev(scales_l)
+    g_idx = stack_and_dev(g_idx_l)
+    sort_indices = stack_and_dev(sort_indices_l)
 
-    zp = torch.empty((0, 0), dtype=dtype, device="cuda", requires_grad=False)
-    opcheck(
-        ops.marlin_gemm_moe,
-        (
-            a,
-            qweight1,
-            sorted_token_ids,
-            topk_weights,
-            topk_ids,
-            scales1,
-            zp,
-            g_idx1,
-            sort_indices1,
-            workspace,
-            quant_type.id,
-            m,
-            2 * n,
-            k,
-            True,
-            e,
-            topk,
-            block_size_m,
-            True,
-            False,
-        ),
+    score = torch.randn((m, e), device="cuda", dtype=dtype)
+    marlin_output = ops.single_marlin_moe(
+        a,
+        qweight,
+        scales,
+        score,
+        topk,
+        renormalize=False,
+        g_idx=g_idx,
+        sort_indices=sort_indices,
+        num_bits=num_bits,
+        is_k_full=is_k_full,
     )
 
+    torch_output = torch_moe_single(a, w_ref.transpose(1, 2), score, topk)
+
+    assert compute_max_diff(marlin_output, torch_output) < 1e-2
+
 
 def test_moe_align_block_size_opcheck():
     num_experts = 4
     block_size = 4
-    topk_ids = torch.randint(0, num_experts, (3, 4), dtype=torch.int32, device="cuda")
+    topk_ids = torch.randint(0,
+                             num_experts, (3, 4),
+                             dtype=torch.int32,
+                             device='cuda')
 
     max_num_tokens_padded = topk_ids.numel() + num_experts * (block_size - 1)
-    sorted_ids = torch.empty(
-        (max_num_tokens_padded,), dtype=torch.int32, device=topk_ids.device
-    )
+    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 = max_num_tokens_padded // block_size
-    expert_ids = torch.empty(
-        (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)
-
-    opcheck(
-        ops.moe_align_block_size,
-        (
-            topk_ids,
-            num_experts,
-            block_size,
-            sorted_ids,
-            expert_ids,
-            num_tokens_post_pad,
-        ),
-    )
+    expert_ids = torch.empty((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)
+
+    opcheck(ops.moe_align_block_size,
+            (topk_ids, num_experts, block_size, sorted_ids, expert_ids,
+             num_tokens_post_pad))
diff --git a/tests/kernels/utils.py b/tests/kernels/utils.py
index c449f6c50cbd46e7212eb68c867d81519f90d32b..134e280092e11ef2e7436009a0449f363f2c7138 100644
--- a/tests/kernels/utils.py
+++ b/tests/kernels/utils.py
@@ -38,7 +38,7 @@ class SiluAndMul(nn.Module):
         return F.silu(x[..., :d]) * x[..., d:]
 
 
-def torch_moe(a, w1, w2, score, topk):
+def torch_moe(a, w1, w2, score, topk, expert_map):
     B, D = a.shape
     a = a.view(B, -1, D).repeat(1, topk, 1).reshape(-1, D)
     out = torch.zeros(B * topk, w2.shape[1], dtype=a.dtype, device=a.device)
@@ -46,15 +46,15 @@ def torch_moe(a, w1, w2, score, topk):
     topk_weight, topk_ids = torch.topk(score, topk)
     topk_weight = topk_weight.view(-1)
     topk_ids = topk_ids.view(-1)
+    if expert_map is not None:
+        topk_ids = expert_map[topk_ids]
     for i in range(w1.shape[0]):
         mask = topk_ids == i
         if mask.sum():
-            out[mask] = SiluAndMul()(a[mask] @ w1[i].transpose(0, 1)) @ w2[i].transpose(
-                0, 1
-            )
-    return (
-        out.view(B, -1, w2.shape[1]) * topk_weight.view(B, -1, 1).to(out.dtype)
-    ).sum(dim=1)
+            out[mask] = SiluAndMul()(
+                a[mask] @ w1[i].transpose(0, 1)) @ w2[i].transpose(0, 1)
+    return (out.view(B, -1, w2.shape[1]) *
+            topk_weight.view(B, -1, 1).to(out.dtype)).sum(dim=1)
 
 
 # Copied/modified from torch._refs.__init__.py
diff --git a/torch-ext/moe/__init__.py b/torch-ext/moe/__init__.py
index cc806778863c03ccb3157343cd6331c1c6ca332c..0d9d8227711def6d51c24b5fd779a3d13681d88c 100644
--- a/torch-ext/moe/__init__.py
+++ b/torch-ext/moe/__init__.py
@@ -1,5 +1,6 @@
 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
@@ -51,24 +52,6 @@ def moe_sum(input: torch.Tensor, output: torch.Tensor):
     ops.moe_sum(input, output)
 
 
-def moe_align_block_size(
-    topk_ids: torch.Tensor,
-    num_experts: int,
-    block_size: int,
-    sorted_token_ids: torch.Tensor,
-    experts_ids: torch.Tensor,
-    num_tokens_post_pad: torch.Tensor,
-) -> None:
-    ops.moe_align_block_size(
-        topk_ids,
-        num_experts,
-        block_size,
-        sorted_token_ids,
-        experts_ids,
-        num_tokens_post_pad,
-    )
-
-
 def topk_softmax(
     topk_weights: torch.Tensor,
     topk_ids: torch.Tensor,
@@ -87,6 +70,7 @@ __all__ = [
     "fused_topk",
     "gptq_marlin_moe_repack",
     "grouped_topk",
+    "layers",
     "moe_align_block_size",
     "moe_sum",
     "per_token_group_quant_fp8",
diff --git a/torch-ext/moe/configs/E=160,N=192,device_name=NVIDIA_A800-SXM4-80GB.json b/torch-ext/moe/configs/E=160,N=192,device_name=NVIDIA_A800-SXM4-80GB.json
new file mode 100644
index 0000000000000000000000000000000000000000..0611620eb33626ac891454637723eabec9aa1183
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/configs/E=256,N=1024,device_name=AMD_Instinct_MI325X,block_shape=[128,128].json b/torch-ext/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/torch-ext/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/torch-ext/moe/configs/E=256,N=1024,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json b/torch-ext/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/torch-ext/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/torch-ext/moe/configs/E=256,N=128,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json b/torch-ext/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/torch-ext/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/torch-ext/moe/configs/E=256,N=128,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8.json b/torch-ext/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/torch-ext/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/torch-ext/moe/configs/E=256,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json b/torch-ext/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/torch-ext/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/torch-ext/moe/configs/E=256,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8.json b/torch-ext/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/torch-ext/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/torch-ext/moe/configs/E=256,N=128,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json b/torch-ext/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/torch-ext/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/torch-ext/moe/configs/E=256,N=128,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128,128].json b/torch-ext/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/torch-ext/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/torch-ext/moe/configs/E=256,N=256,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json b/torch-ext/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/torch-ext/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/torch-ext/moe/configs/E=256,N=256,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json b/torch-ext/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/torch-ext/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/torch-ext/moe/configs/E=256,N=256,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json b/torch-ext/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/torch-ext/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/torch-ext/moe/configs/E=256,N=256,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json b/torch-ext/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/torch-ext/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/torch-ext/moe/configs/E=256,N=256,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128,128].json b/torch-ext/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/torch-ext/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/torch-ext/moe/configs/E=256,N=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json b/torch-ext/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/torch-ext/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/torch-ext/moe/configs/E=256,N=256,device_name=NVIDIA_L20,dtype=fp8_w8a8,block_shape=[128,128].json b/torch-ext/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/torch-ext/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/torch-ext/moe/configs/E=256,N=512,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json b/torch-ext/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/torch-ext/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/torch-ext/moe/configs/E=256,N=64,device_name=NVIDIA_A800-SXM4-80GB.json b/torch-ext/moe/configs/E=256,N=64,device_name=NVIDIA_A800-SXM4-80GB.json
new file mode 100644
index 0000000000000000000000000000000000000000..21f60229ff875ca6dc0cd0a3cb4fa51a845bf881
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/configs/E=60,N=1408,device_name=AMD_Instinct_MI300X.json b/torch-ext/moe/configs/E=60,N=1408,device_name=AMD_Instinct_MI300X.json
new file mode 100644
index 0000000000000000000000000000000000000000..d09508b31729ddbafc8b5e1dfa88343c61e83ca9
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/configs/E=60,N=176,device_name=AMD_Instinct_MI300X.json b/torch-ext/moe/configs/E=60,N=176,device_name=AMD_Instinct_MI300X.json
new file mode 100644
index 0000000000000000000000000000000000000000..746463af4d56315af3c31348c2253e43c926553e
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/configs/E=60,N=352,device_name=AMD_Instinct_MI300X.json b/torch-ext/moe/configs/E=60,N=352,device_name=AMD_Instinct_MI300X.json
new file mode 100644
index 0000000000000000000000000000000000000000..bbdb9ad096451d6c5b9d95fae4ed02ce37715002
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/configs/E=60,N=704,device_name=AMD_Instinct_MI300X.json b/torch-ext/moe/configs/E=60,N=704,device_name=AMD_Instinct_MI300X.json
new file mode 100644
index 0000000000000000000000000000000000000000..43584b1eb6b65f02fc919490b0ff4b9569de58c4
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/configs/E=64,N=1280,device_name=NVIDIA_A800-SXM4-80GB.json b/torch-ext/moe/configs/E=64,N=1280,device_name=NVIDIA_A800-SXM4-80GB.json
new file mode 100644
index 0000000000000000000000000000000000000000..39a9912fa4bddc48de38d5a8b4fadd71326485d2
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/configs/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/torch-ext/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/torch-ext/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/torch-ext/moe/configs/E=64,N=1280,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/torch-ext/moe/configs/E=64,N=1280,device_name=NVIDIA_H200,dtype=fp8_w8a8.json
new file mode 100644
index 0000000000000000000000000000000000000000..c17a4ec346915af69f10f65410450cbd66614810
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/configs/E=64,N=1280,device_name=NVIDIA_H200.json b/torch-ext/moe/configs/E=64,N=1280,device_name=NVIDIA_H200.json
new file mode 100644
index 0000000000000000000000000000000000000000..170ae7f3fff1d09961d368853dcfc2f24c1b878c
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/configs/E=64,N=2560,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/torch-ext/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/torch-ext/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/torch-ext/moe/configs/E=64,N=2560,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/torch-ext/moe/configs/E=64,N=2560,device_name=NVIDIA_H200,dtype=fp8_w8a8.json
new file mode 100644
index 0000000000000000000000000000000000000000..9ad5b316750051e4a5dc39b195f6a29eeaa882f0
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/configs/E=64,N=2560,device_name=NVIDIA_H200.json b/torch-ext/moe/configs/E=64,N=2560,device_name=NVIDIA_H200.json
new file mode 100644
index 0000000000000000000000000000000000000000..2883dfd11e7f3a7db429ae0aaeb98d554f271cdf
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/configs/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/torch-ext/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/torch-ext/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/torch-ext/moe/configs/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3.json b/torch-ext/moe/configs/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3.json
new file mode 100644
index 0000000000000000000000000000000000000000..2fc18a5e43d298ddabc57dc9f357296890fb7e29
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/configs/E=64,N=320,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/torch-ext/moe/configs/E=64,N=320,device_name=NVIDIA_H200,dtype=fp8_w8a8.json
new file mode 100644
index 0000000000000000000000000000000000000000..be8d4a7fd23d994ac838d45bf249e39764552ccf
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/configs/E=64,N=320,device_name=NVIDIA_H200.json b/torch-ext/moe/configs/E=64,N=320,device_name=NVIDIA_H200.json
new file mode 100644
index 0000000000000000000000000000000000000000..71fdd88643c6ff515808fdc2d67e9fda3ac62ae9
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/configs/E=64,N=640,device_name=NVIDIA_A800-SXM4-80GB.json b/torch-ext/moe/configs/E=64,N=640,device_name=NVIDIA_A800-SXM4-80GB.json
new file mode 100644
index 0000000000000000000000000000000000000000..c02de2f628b716456373978f600f9fc8c4b98d24
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/configs/E=64,N=640,device_name=NVIDIA_GeForce_RTX_4090,dtype=fp8_w8a8.json b/torch-ext/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/torch-ext/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/torch-ext/moe/configs/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/torch-ext/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/torch-ext/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/torch-ext/moe/configs/E=64,N=640,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/torch-ext/moe/configs/E=64,N=640,device_name=NVIDIA_H200,dtype=fp8_w8a8.json
new file mode 100644
index 0000000000000000000000000000000000000000..21b72557e365d01011f48a046915765e67dbd13f
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/configs/E=64,N=640,device_name=NVIDIA_H200.json b/torch-ext/moe/configs/E=64,N=640,device_name=NVIDIA_H200.json
new file mode 100644
index 0000000000000000000000000000000000000000..eaf32f6d76c0a6aa50a5fbc2f055a5e7c9e10b66
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/torch-ext/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/torch-ext/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/torch-ext/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X.json b/torch-ext/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X.json
index 6a976788f9b10af19ebcfe582a69cbc627f9457b..4bf775347eccb3e76e20ea0018d71343f9848313 100644
--- a/torch-ext/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X.json
+++ b/torch-ext/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X.json
@@ -1,123 +1,123 @@
 {
     "1": {
         "BLOCK_SIZE_M": 16,
-        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_N": 16,
         "BLOCK_SIZE_K": 256,
         "GROUP_SIZE_M": 1,
         "num_warps": 2,
-        "num_stages": 0,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
-        "kpack": 1
+        "kpack": 2
     },
     "2": {
         "BLOCK_SIZE_M": 16,
         "BLOCK_SIZE_N": 16,
-        "BLOCK_SIZE_K": 128,
+        "BLOCK_SIZE_K": 256,
         "GROUP_SIZE_M": 1,
-        "num_warps": 2,
-        "num_stages": 0,
+        "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": 256,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
         "GROUP_SIZE_M": 1,
-        "num_warps": 2,
-        "num_stages": 0,
+        "num_warps": 1,
+        "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,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
         "GROUP_SIZE_M": 1,
-        "num_warps": 1,
-        "num_stages": 0,
+        "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,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
         "GROUP_SIZE_M": 1,
-        "num_warps": 4,
-        "num_stages": 0,
+        "num_warps": 2,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
         "kpack": 2
     },
     "24": {
         "BLOCK_SIZE_M": 16,
-        "BLOCK_SIZE_N": 32,
-        "BLOCK_SIZE_K": 64,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 256,
         "GROUP_SIZE_M": 1,
-        "num_warps": 1,
-        "num_stages": 0,
+        "num_warps": 2,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
         "kpack": 2
     },
     "32": {
         "BLOCK_SIZE_M": 16,
-        "BLOCK_SIZE_N": 16,
-        "BLOCK_SIZE_K": 128,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
         "GROUP_SIZE_M": 4,
         "num_warps": 2,
-        "num_stages": 0,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
-        "kpack": 1
+        "kpack": 2
     },
     "48": {
         "BLOCK_SIZE_M": 16,
-        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_N": 64,
         "BLOCK_SIZE_K": 128,
         "GROUP_SIZE_M": 4,
-        "num_warps": 2,
-        "num_stages": 0,
+        "num_warps": 4,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
-        "kpack": 2
+        "kpack": 1
     },
     "64": {
         "BLOCK_SIZE_M": 32,
         "BLOCK_SIZE_N": 64,
         "BLOCK_SIZE_K": 128,
         "GROUP_SIZE_M": 4,
-        "num_warps": 8,
-        "num_stages": 0,
+        "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": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
         "GROUP_SIZE_M": 4,
-        "num_warps": 4,
-        "num_stages": 0,
+        "num_warps": 8,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
-        "kpack": 2
+        "kpack": 1
     },
     "128": {
         "BLOCK_SIZE_M": 64,
         "BLOCK_SIZE_N": 64,
-        "BLOCK_SIZE_K": 64,
+        "BLOCK_SIZE_K": 128,
         "GROUP_SIZE_M": 4,
-        "num_warps": 8,
-        "num_stages": 0,
+        "num_warps": 4,
+        "num_stages": 2,
         "waves_per_eu": 0,
-        "matrix_instr_nonkdim": 16,
+        "matrix_instr_nonkdim": 32,
         "kpack": 2
     },
     "256": {
@@ -126,10 +126,10 @@
         "BLOCK_SIZE_K": 64,
         "GROUP_SIZE_M": 4,
         "num_warps": 8,
-        "num_stages": 0,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
-        "kpack": 1
+        "kpack": 2
     },
     "512": {
         "BLOCK_SIZE_M": 128,
@@ -137,7 +137,7 @@
         "BLOCK_SIZE_K": 64,
         "GROUP_SIZE_M": 4,
         "num_warps": 8,
-        "num_stages": 0,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
         "kpack": 2
@@ -148,9 +148,9 @@
         "BLOCK_SIZE_K": 64,
         "GROUP_SIZE_M": 1,
         "num_warps": 8,
-        "num_stages": 0,
+        "num_stages": 2,
         "waves_per_eu": 0,
-        "matrix_instr_nonkdim": 32,
+        "matrix_instr_nonkdim": 16,
         "kpack": 2
     },
     "1536": {
@@ -159,7 +159,7 @@
         "BLOCK_SIZE_K": 64,
         "GROUP_SIZE_M": 1,
         "num_warps": 8,
-        "num_stages": 0,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
         "kpack": 2
@@ -170,7 +170,7 @@
         "BLOCK_SIZE_K": 64,
         "GROUP_SIZE_M": 1,
         "num_warps": 8,
-        "num_stages": 0,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
         "kpack": 2
@@ -181,10 +181,10 @@
         "BLOCK_SIZE_K": 64,
         "GROUP_SIZE_M": 1,
         "num_warps": 8,
-        "num_stages": 0,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
-        "kpack": 1
+        "kpack": 2
     },
     "4096": {
         "BLOCK_SIZE_M": 128,
@@ -192,9 +192,9 @@
         "BLOCK_SIZE_K": 64,
         "GROUP_SIZE_M": 1,
         "num_warps": 8,
-        "num_stages": 0,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
-        "kpack": 1
+        "kpack": 2
     }
 }
diff --git a/torch-ext/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json b/torch-ext/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/torch-ext/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/torch-ext/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI325X.json b/torch-ext/moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI325X.json
new file mode 100644
index 0000000000000000000000000000000000000000..3918c93b160a39e55153cbf6ff634d59e724d021
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/configs/E=8,N=14336,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/torch-ext/moe/configs/E=8,N=14336,device_name=NVIDIA_H200,dtype=fp8_w8a8.json
new file mode 100644
index 0000000000000000000000000000000000000000..841044a4fc6e2da81f18b2765fa30b9de7092b3d
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/configs/E=8,N=14336,device_name=NVIDIA_H200.json b/torch-ext/moe/configs/E=8,N=14336,device_name=NVIDIA_H200.json
new file mode 100644
index 0000000000000000000000000000000000000000..59be497fc428755f010e490277f4b945d6ed0195
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/torch-ext/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/torch-ext/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/torch-ext/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI300X.json b/torch-ext/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI300X.json
new file mode 100644
index 0000000000000000000000000000000000000000..d6ad63509f1573d25fbdc370b93bc39a0d00b51a
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json b/torch-ext/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/torch-ext/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/torch-ext/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI325X.json b/torch-ext/moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI325X.json
new file mode 100644
index 0000000000000000000000000000000000000000..d766fc062ddcce10d38bdc68ab1f700d193b8ba7
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/torch-ext/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/torch-ext/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/torch-ext/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI300X.json b/torch-ext/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI300X.json
index 0a46390b2e31bba6a7c3ab2c9f6c8de6004857bb..1b46cb5716514f290c6b4bc3a18dba490c799272 100644
--- a/torch-ext/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI300X.json
+++ b/torch-ext/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI300X.json
@@ -1,11 +1,11 @@
 {
     "1": {
         "BLOCK_SIZE_M": 16,
-        "BLOCK_SIZE_N": 32,
-        "BLOCK_SIZE_K": 256,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 128,
         "GROUP_SIZE_M": 1,
-        "num_warps": 2,
-        "num_stages": 0,
+        "num_warps": 4,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
         "kpack": 2
@@ -16,95 +16,95 @@
         "BLOCK_SIZE_K": 128,
         "GROUP_SIZE_M": 1,
         "num_warps": 4,
-        "num_stages": 0,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
-        "kpack": 1
+        "kpack": 2
     },
     "4": {
         "BLOCK_SIZE_M": 16,
-        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_N": 128,
         "BLOCK_SIZE_K": 128,
         "GROUP_SIZE_M": 1,
-        "num_warps": 4,
-        "num_stages": 0,
+        "num_warps": 8,
+        "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,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
         "GROUP_SIZE_M": 1,
-        "num_warps": 2,
-        "num_stages": 0,
+        "num_warps": 4,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
-        "kpack": 2
+        "kpack": 1
     },
     "16": {
-        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_M": 16,
         "BLOCK_SIZE_N": 64,
-        "BLOCK_SIZE_K": 64,
+        "BLOCK_SIZE_K": 128,
         "GROUP_SIZE_M": 1,
-        "num_warps": 8,
-        "num_stages": 0,
+        "num_warps": 4,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
-        "kpack": 1
+        "kpack": 2
     },
     "24": {
         "BLOCK_SIZE_M": 16,
         "BLOCK_SIZE_N": 64,
-        "BLOCK_SIZE_K": 64,
+        "BLOCK_SIZE_K": 128,
         "GROUP_SIZE_M": 1,
         "num_warps": 4,
-        "num_stages": 0,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
-        "kpack": 1
+        "kpack": 2
     },
     "32": {
         "BLOCK_SIZE_M": 16,
-        "BLOCK_SIZE_N": 16,
-        "BLOCK_SIZE_K": 256,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
         "GROUP_SIZE_M": 4,
-        "num_warps": 2,
-        "num_stages": 0,
+        "num_warps": 4,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
-        "kpack": 2
+        "kpack": 1
     },
     "48": {
         "BLOCK_SIZE_M": 16,
-        "BLOCK_SIZE_N": 128,
-        "BLOCK_SIZE_K": 64,
-        "GROUP_SIZE_M": 1,
-        "num_warps": 8,
-        "num_stages": 0,
+        "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": 16,
+        "BLOCK_SIZE_M": 32,
         "BLOCK_SIZE_N": 64,
-        "BLOCK_SIZE_K": 64,
+        "BLOCK_SIZE_K": 128,
         "GROUP_SIZE_M": 1,
-        "num_warps": 2,
-        "num_stages": 0,
+        "num_warps": 4,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
-        "kpack": 1
+        "kpack": 2
     },
     "96": {
         "BLOCK_SIZE_M": 32,
-        "BLOCK_SIZE_N": 32,
-        "BLOCK_SIZE_K": 256,
-        "GROUP_SIZE_M": 4,
-        "num_warps": 4,
-        "num_stages": 0,
+        "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
@@ -112,24 +112,24 @@
     "128": {
         "BLOCK_SIZE_M": 64,
         "BLOCK_SIZE_N": 64,
-        "BLOCK_SIZE_K": 64,
+        "BLOCK_SIZE_K": 128,
         "GROUP_SIZE_M": 4,
         "num_warps": 8,
-        "num_stages": 0,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
-        "kpack": 1
+        "kpack": 2
     },
     "256": {
-        "BLOCK_SIZE_M": 128,
-        "BLOCK_SIZE_N": 128,
-        "BLOCK_SIZE_K": 64,
-        "GROUP_SIZE_M": 4,
+        "BLOCK_SIZE_M": 32,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
         "num_warps": 8,
-        "num_stages": 0,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
-        "kpack": 1
+        "kpack": 2
     },
     "512": {
         "BLOCK_SIZE_M": 64,
@@ -137,7 +137,7 @@
         "BLOCK_SIZE_K": 128,
         "GROUP_SIZE_M": 1,
         "num_warps": 8,
-        "num_stages": 0,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
         "kpack": 2
@@ -148,10 +148,10 @@
         "BLOCK_SIZE_K": 64,
         "GROUP_SIZE_M": 1,
         "num_warps": 8,
-        "num_stages": 0,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
-        "kpack": 1
+        "kpack": 2
     },
     "1536": {
         "BLOCK_SIZE_M": 128,
@@ -159,10 +159,10 @@
         "BLOCK_SIZE_K": 64,
         "GROUP_SIZE_M": 1,
         "num_warps": 8,
-        "num_stages": 0,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
-        "kpack": 1
+        "kpack": 2
     },
     "2048": {
         "BLOCK_SIZE_M": 128,
@@ -170,7 +170,7 @@
         "BLOCK_SIZE_K": 64,
         "GROUP_SIZE_M": 1,
         "num_warps": 8,
-        "num_stages": 0,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
         "kpack": 2
@@ -181,10 +181,10 @@
         "BLOCK_SIZE_K": 64,
         "GROUP_SIZE_M": 1,
         "num_warps": 8,
-        "num_stages": 0,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
-        "kpack": 1
+        "kpack": 2
     },
     "4096": {
         "BLOCK_SIZE_M": 128,
@@ -192,9 +192,9 @@
         "BLOCK_SIZE_K": 64,
         "GROUP_SIZE_M": 1,
         "num_warps": 8,
-        "num_stages": 0,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
-        "kpack": 1
+        "kpack": 2
     }
 }
diff --git a/torch-ext/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json b/torch-ext/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/torch-ext/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/torch-ext/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI325X.json b/torch-ext/moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI325X.json
new file mode 100644
index 0000000000000000000000000000000000000000..ffc1b23ea90ddefd24b4b98a222e67e84442cf9c
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/configs/E=8,N=1792,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/torch-ext/moe/configs/E=8,N=1792,device_name=NVIDIA_H200,dtype=fp8_w8a8.json
new file mode 100644
index 0000000000000000000000000000000000000000..e4110a5d2e70f4d2698d4c28b1668f8e4a48688a
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/configs/E=8,N=1792,device_name=NVIDIA_H200.json b/torch-ext/moe/configs/E=8,N=1792,device_name=NVIDIA_H200.json
new file mode 100644
index 0000000000000000000000000000000000000000..0883ef40582ea4f1b08c629edcd6dd9beaaf2dad
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/torch-ext/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/torch-ext/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/torch-ext/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI300X.json b/torch-ext/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI300X.json
new file mode 100644
index 0000000000000000000000000000000000000000..811c77ab4109319c9f079e414bd4f7d2cf4ac9f1
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json b/torch-ext/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/torch-ext/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/torch-ext/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI325X.json b/torch-ext/moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI325X.json
new file mode 100644
index 0000000000000000000000000000000000000000..fc31215cbae8afeee27d923e176a25daf1f2b43d
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/configs/E=8,N=2048,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/torch-ext/moe/configs/E=8,N=2048,device_name=NVIDIA_H200,dtype=fp8_w8a8.json
new file mode 100644
index 0000000000000000000000000000000000000000..1a0aa331933291f0c32b2cf26672771b5c7b4196
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/configs/E=8,N=2048,device_name=NVIDIA_H200.json b/torch-ext/moe/configs/E=8,N=2048,device_name=NVIDIA_H200.json
new file mode 100644
index 0000000000000000000000000000000000000000..9952be6ba4abe434e2924e9505e834de4ce0528e
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/torch-ext/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/torch-ext/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/torch-ext/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X.json b/torch-ext/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X.json
index 91011e64c7de4505e9bb462bc70e6a3e7affa878..5a3f415d54142383fa7677d8192f7a4c4df66b2e 100644
--- a/torch-ext/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X.json
+++ b/torch-ext/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X.json
@@ -1,91 +1,91 @@
 {
     "1": {
         "BLOCK_SIZE_M": 16,
-        "BLOCK_SIZE_N": 16,
-        "BLOCK_SIZE_K": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
         "GROUP_SIZE_M": 1,
-        "num_warps": 2,
-        "num_stages": 0,
+        "num_warps": 4,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
-        "kpack": 1
+        "kpack": 2
     },
     "2": {
         "BLOCK_SIZE_M": 16,
-        "BLOCK_SIZE_N": 16,
-        "BLOCK_SIZE_K": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
         "GROUP_SIZE_M": 1,
-        "num_warps": 2,
-        "num_stages": 0,
+        "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": 256,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
         "GROUP_SIZE_M": 1,
-        "num_warps": 2,
-        "num_stages": 0,
+        "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_N": 16,
         "BLOCK_SIZE_K": 256,
         "GROUP_SIZE_M": 1,
         "num_warps": 2,
-        "num_stages": 0,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
         "kpack": 2
     },
     "16": {
         "BLOCK_SIZE_M": 16,
-        "BLOCK_SIZE_N": 32,
-        "BLOCK_SIZE_K": 256,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
         "GROUP_SIZE_M": 1,
         "num_warps": 2,
-        "num_stages": 0,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
-        "kpack": 2
+        "kpack": 1
     },
     "24": {
         "BLOCK_SIZE_M": 16,
-        "BLOCK_SIZE_N": 64,
-        "BLOCK_SIZE_K": 64,
+        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_K": 256,
         "GROUP_SIZE_M": 1,
-        "num_warps": 4,
-        "num_stages": 0,
+        "num_warps": 2,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
-        "kpack": 1
+        "kpack": 2
     },
     "32": {
         "BLOCK_SIZE_M": 16,
-        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_N": 32,
         "BLOCK_SIZE_K": 256,
         "GROUP_SIZE_M": 4,
         "num_warps": 2,
-        "num_stages": 0,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
-        "kpack": 2
+        "kpack": 1
     },
     "48": {
-        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_M": 32,
         "BLOCK_SIZE_N": 32,
         "BLOCK_SIZE_K": 256,
         "GROUP_SIZE_M": 1,
         "num_warps": 2,
-        "num_stages": 0,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
-        "kpack": 2
+        "kpack": 1
     },
     "64": {
         "BLOCK_SIZE_M": 32,
@@ -93,7 +93,7 @@
         "BLOCK_SIZE_K": 256,
         "GROUP_SIZE_M": 4,
         "num_warps": 4,
-        "num_stages": 0,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
         "kpack": 2
@@ -101,43 +101,43 @@
     "96": {
         "BLOCK_SIZE_M": 32,
         "BLOCK_SIZE_N": 32,
-        "BLOCK_SIZE_K": 128,
+        "BLOCK_SIZE_K": 256,
         "GROUP_SIZE_M": 4,
-        "num_warps": 4,
-        "num_stages": 0,
+        "num_warps": 2,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
-        "kpack": 1
+        "kpack": 2
     },
     "128": {
         "BLOCK_SIZE_M": 64,
         "BLOCK_SIZE_N": 64,
-        "BLOCK_SIZE_K": 128,
-        "GROUP_SIZE_M": 4,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
         "num_warps": 8,
-        "num_stages": 0,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
-        "kpack": 1
+        "kpack": 2
     },
     "256": {
-        "BLOCK_SIZE_M": 128,
-        "BLOCK_SIZE_N": 128,
-        "BLOCK_SIZE_K": 64,
-        "GROUP_SIZE_M": 4,
+        "BLOCK_SIZE_M": 64,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 128,
+        "GROUP_SIZE_M": 1,
         "num_warps": 8,
-        "num_stages": 0,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
-        "kpack": 1
+        "kpack": 2
     },
     "512": {
         "BLOCK_SIZE_M": 64,
         "BLOCK_SIZE_N": 128,
-        "BLOCK_SIZE_K": 64,
+        "BLOCK_SIZE_K": 128,
         "GROUP_SIZE_M": 1,
         "num_warps": 8,
-        "num_stages": 0,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 32,
         "kpack": 2
@@ -148,10 +148,10 @@
         "BLOCK_SIZE_K": 64,
         "GROUP_SIZE_M": 1,
         "num_warps": 8,
-        "num_stages": 0,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
-        "kpack": 1
+        "kpack": 2
     },
     "1536": {
         "BLOCK_SIZE_M": 128,
@@ -159,7 +159,7 @@
         "BLOCK_SIZE_K": 64,
         "GROUP_SIZE_M": 1,
         "num_warps": 8,
-        "num_stages": 0,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
         "kpack": 2
@@ -170,10 +170,10 @@
         "BLOCK_SIZE_K": 64,
         "GROUP_SIZE_M": 1,
         "num_warps": 8,
-        "num_stages": 0,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
-        "kpack": 1
+        "kpack": 2
     },
     "3072": {
         "BLOCK_SIZE_M": 128,
@@ -181,7 +181,7 @@
         "BLOCK_SIZE_K": 64,
         "GROUP_SIZE_M": 1,
         "num_warps": 8,
-        "num_stages": 0,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
         "kpack": 2
@@ -192,9 +192,9 @@
         "BLOCK_SIZE_K": 64,
         "GROUP_SIZE_M": 1,
         "num_warps": 8,
-        "num_stages": 0,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
-        "kpack": 1
+        "kpack": 2
     }
 }
diff --git a/torch-ext/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json b/torch-ext/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/torch-ext/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/torch-ext/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI325X.json b/torch-ext/moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI325X.json
new file mode 100644
index 0000000000000000000000000000000000000000..de9d0aba75a7ed45fdfd92c123736f1e981a06eb
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/configs/E=8,N=3584,device_name=NVIDIA_GeForce_RTX_4090,dtype=fp8_w8a8.json b/torch-ext/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/torch-ext/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/torch-ext/moe/configs/E=8,N=3584,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/torch-ext/moe/configs/E=8,N=3584,device_name=NVIDIA_H200,dtype=fp8_w8a8.json
new file mode 100644
index 0000000000000000000000000000000000000000..e6f753cdba35b55e8d021597dc4911bab27caa8d
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/configs/E=8,N=3584,device_name=NVIDIA_H200.json b/torch-ext/moe/configs/E=8,N=3584,device_name=NVIDIA_H200.json
new file mode 100644
index 0000000000000000000000000000000000000000..53f3394693f060daebff9960a95ed3753e8522c9
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/torch-ext/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/torch-ext/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/torch-ext/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI300X.json b/torch-ext/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI300X.json
new file mode 100644
index 0000000000000000000000000000000000000000..a64d06c6d1724f736c5081a8b947fa2e6127701d
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json b/torch-ext/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/torch-ext/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/torch-ext/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI325X.json b/torch-ext/moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI325X.json
new file mode 100644
index 0000000000000000000000000000000000000000..c7db6c0cbd3faf615074fc8055ff744e528dd204
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/configs/E=8,N=4096,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/torch-ext/moe/configs/E=8,N=4096,device_name=NVIDIA_H200,dtype=fp8_w8a8.json
new file mode 100644
index 0000000000000000000000000000000000000000..4dd475c02a19b380efbe8f59945a0f2a41acda11
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/configs/E=8,N=4096,device_name=NVIDIA_H200.json b/torch-ext/moe/configs/E=8,N=4096,device_name=NVIDIA_H200.json
new file mode 100644
index 0000000000000000000000000000000000000000..2ed15f30fe603c080a2e8a32b664400a50d2153e
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/torch-ext/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/torch-ext/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/torch-ext/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X.json b/torch-ext/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X.json
index f807d4a5abaed9dd686df26837f2dd9f6161300f..8d7b780271854c0fbae29e92cfd3cce508ca029e 100644
--- a/torch-ext/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X.json
+++ b/torch-ext/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X.json
@@ -1,66 +1,66 @@
 {
     "1": {
         "BLOCK_SIZE_M": 16,
-        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_N": 32,
         "BLOCK_SIZE_K": 256,
         "GROUP_SIZE_M": 1,
         "num_warps": 2,
-        "num_stages": 0,
+        "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": 32,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 256,
         "GROUP_SIZE_M": 1,
-        "num_warps": 4,
-        "num_stages": 0,
+        "num_warps": 2,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
-        "kpack": 1
+        "kpack": 2
     },
     "4": {
         "BLOCK_SIZE_M": 16,
-        "BLOCK_SIZE_N": 32,
-        "BLOCK_SIZE_K": 128,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 256,
         "GROUP_SIZE_M": 1,
-        "num_warps": 4,
-        "num_stages": 0,
+        "num_warps": 2,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
-        "kpack": 1
+        "kpack": 2
     },
     "8": {
         "BLOCK_SIZE_M": 16,
-        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_N": 16,
         "BLOCK_SIZE_K": 256,
         "GROUP_SIZE_M": 1,
-        "num_warps": 2,
-        "num_stages": 0,
+        "num_warps": 1,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
-        "kpack": 1
+        "kpack": 2
     },
     "16": {
         "BLOCK_SIZE_M": 16,
-        "BLOCK_SIZE_N": 16,
-        "BLOCK_SIZE_K": 256,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 64,
         "GROUP_SIZE_M": 1,
-        "num_warps": 4,
-        "num_stages": 0,
+        "num_warps": 2,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
-        "kpack": 2
+        "kpack": 1
     },
     "24": {
-        "BLOCK_SIZE_M": 32,
-        "BLOCK_SIZE_N": 32,
-        "BLOCK_SIZE_K": 128,
+        "BLOCK_SIZE_M": 16,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 256,
         "GROUP_SIZE_M": 1,
-        "num_warps": 8,
-        "num_stages": 0,
+        "num_warps": 1,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
         "kpack": 1
@@ -68,43 +68,43 @@
     "32": {
         "BLOCK_SIZE_M": 16,
         "BLOCK_SIZE_N": 32,
-        "BLOCK_SIZE_K": 128,
+        "BLOCK_SIZE_K": 256,
         "GROUP_SIZE_M": 4,
         "num_warps": 2,
-        "num_stages": 0,
+        "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": 128,
-        "GROUP_SIZE_M": 4,
+        "BLOCK_SIZE_N": 16,
+        "BLOCK_SIZE_K": 256,
+        "GROUP_SIZE_M": 1,
         "num_warps": 2,
-        "num_stages": 0,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
-        "kpack": 1
+        "kpack": 2
     },
     "64": {
         "BLOCK_SIZE_M": 32,
-        "BLOCK_SIZE_N": 32,
+        "BLOCK_SIZE_N": 64,
         "BLOCK_SIZE_K": 128,
         "GROUP_SIZE_M": 4,
         "num_warps": 4,
-        "num_stages": 0,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
-        "kpack": 2
+        "kpack": 1
     },
     "96": {
         "BLOCK_SIZE_M": 32,
-        "BLOCK_SIZE_N": 32,
-        "BLOCK_SIZE_K": 128,
+        "BLOCK_SIZE_N": 64,
+        "BLOCK_SIZE_K": 256,
         "GROUP_SIZE_M": 4,
-        "num_warps": 4,
-        "num_stages": 0,
+        "num_warps": 8,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
         "kpack": 2
@@ -112,23 +112,23 @@
     "128": {
         "BLOCK_SIZE_M": 64,
         "BLOCK_SIZE_N": 64,
-        "BLOCK_SIZE_K": 64,
+        "BLOCK_SIZE_K": 128,
         "GROUP_SIZE_M": 4,
         "num_warps": 8,
-        "num_stages": 0,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
-        "kpack": 1
+        "kpack": 2
     },
     "256": {
         "BLOCK_SIZE_M": 128,
         "BLOCK_SIZE_N": 128,
-        "BLOCK_SIZE_K": 64,
+        "BLOCK_SIZE_K": 128,
         "GROUP_SIZE_M": 4,
         "num_warps": 8,
-        "num_stages": 0,
+        "num_stages": 2,
         "waves_per_eu": 0,
-        "matrix_instr_nonkdim": 32,
+        "matrix_instr_nonkdim": 16,
         "kpack": 2
     },
     "512": {
@@ -137,10 +137,10 @@
         "BLOCK_SIZE_K": 64,
         "GROUP_SIZE_M": 1,
         "num_warps": 8,
-        "num_stages": 0,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
-        "kpack": 1
+        "kpack": 2
     },
     "1024": {
         "BLOCK_SIZE_M": 128,
@@ -148,10 +148,10 @@
         "BLOCK_SIZE_K": 64,
         "GROUP_SIZE_M": 1,
         "num_warps": 8,
-        "num_stages": 0,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
-        "kpack": 1
+        "kpack": 2
     },
     "1536": {
         "BLOCK_SIZE_M": 128,
@@ -159,7 +159,7 @@
         "BLOCK_SIZE_K": 64,
         "GROUP_SIZE_M": 1,
         "num_warps": 8,
-        "num_stages": 0,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
         "kpack": 2
@@ -170,10 +170,10 @@
         "BLOCK_SIZE_K": 64,
         "GROUP_SIZE_M": 1,
         "num_warps": 8,
-        "num_stages": 0,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
-        "kpack": 1
+        "kpack": 2
     },
     "3072": {
         "BLOCK_SIZE_M": 128,
@@ -181,20 +181,20 @@
         "BLOCK_SIZE_K": 64,
         "GROUP_SIZE_M": 1,
         "num_warps": 8,
-        "num_stages": 0,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
         "kpack": 2
     },
     "4096": {
-        "BLOCK_SIZE_M": 128,
+        "BLOCK_SIZE_M": 256,
         "BLOCK_SIZE_N": 128,
         "BLOCK_SIZE_K": 64,
         "GROUP_SIZE_M": 1,
         "num_warps": 8,
-        "num_stages": 0,
+        "num_stages": 2,
         "waves_per_eu": 0,
         "matrix_instr_nonkdim": 16,
-        "kpack": 1
+        "kpack": 2
     }
 }
diff --git a/torch-ext/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json b/torch-ext/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/torch-ext/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/torch-ext/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI325X.json b/torch-ext/moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI325X.json
new file mode 100644
index 0000000000000000000000000000000000000000..3a3268cc17a84d88f0a67e20e266491ad45d1777
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/configs/E=8,N=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/torch-ext/moe/configs/E=8,N=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8.json
new file mode 100644
index 0000000000000000000000000000000000000000..eb817268d41208a661d2ab9526c5108d36d44085
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/configs/E=8,N=7168,device_name=NVIDIA_H200.json b/torch-ext/moe/configs/E=8,N=7168,device_name=NVIDIA_H200.json
new file mode 100644
index 0000000000000000000000000000000000000000..0c7062aea6c4e920cdfdc9e13fc2c0fa38aceb0e
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json b/torch-ext/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/torch-ext/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/torch-ext/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI300X.json b/torch-ext/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI300X.json
new file mode 100644
index 0000000000000000000000000000000000000000..cf66868e9d57a541fe10fa9fff682f7a7c6dac97
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json b/torch-ext/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/torch-ext/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/torch-ext/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI325X.json b/torch-ext/moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI325X.json
new file mode 100644
index 0000000000000000000000000000000000000000..da477b1fb15eb2064cbff78b5e47ae9c54dae7a5
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/configs/E=8,N=8192,device_name=NVIDIA_H200,dtype=fp8_w8a8.json b/torch-ext/moe/configs/E=8,N=8192,device_name=NVIDIA_H200,dtype=fp8_w8a8.json
new file mode 100644
index 0000000000000000000000000000000000000000..96cbc111c7fffe8b0a67c22711f9a5c8f28d8b61
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/configs/README b/torch-ext/moe/configs/README
index 45d40cbfb1a2ef58a14e2fb49b6ff762c12ac77b..787bd061166468b3f213eced25aa963774fbd090 100644
--- a/torch-ext/moe/configs/README
+++ b/torch-ext/moe/configs/README
@@ -8,3 +8,6 @@ the JSON file contains a mapping from M (batch size) to the chosen configuration
 The example configurations provided are for the Mixtral model for TP2 on H100
 and TP4 on A100. Mixtral has intermediate size N = 14336, i.e. for TP2 we have
 N = 7168 and for TP4 we have N = 3584.
+
+Please feel free to tune the configurations using scripts in `benchmarks/kernels/benchmark_moe.py`
+Some of the configurations files are copied from the SGLang repository. Thank you!
diff --git a/torch-ext/moe/fused_moe.py b/torch-ext/moe/fused_moe.py
index af2d798cbe5d7c3c1760ce79f717ab5f6d7700ba..b0d3625037e9509140e8fefae153357606da4325 100644
--- a/torch-ext/moe/fused_moe.py
+++ b/torch-ext/moe/fused_moe.py
@@ -14,7 +14,9 @@ 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__)
 
@@ -25,6 +27,26 @@ VLLM_ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON: bool = bool(
 )
 
 
+@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
@@ -125,14 +147,31 @@ def fused_moe_kernel_gptq_awq(
     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
     )
 
-    off_experts = tl.load(expert_ids_ptr + pid_m).to(tl.int64)
-
     if use_int4_w4a16:
         b_ptrs = (
             b_ptr
@@ -339,13 +378,31 @@ def fused_moe_kernel(
     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
     )
 
-    off_experts = tl.load(expert_ids_ptr + pid_m).to(tl.int64)
     b_ptrs = (
         b_ptr
         + off_experts * stride_be
@@ -374,7 +431,6 @@ def fused_moe_kernel(
     # 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.
@@ -425,222 +481,6 @@ def fused_moe_kernel(
     tl.store(c_ptrs, accumulator, mask=c_mask)
 
 
-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
-) -> 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.
-
-    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)
-    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 = torch.empty(
-        (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:
-            moe_align_block_size_triton(
-                topk_ids,
-                num_experts,
-                block_size,
-                sorted_ids,
-                expert_ids,
-                num_tokens_post_pad,
-            )
-        else:
-            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,
-        )
-    return sorted_ids, expert_ids, num_tokens_post_pad
-
-
 def invoke_fused_moe_kernel(
     A: torch.Tensor,
     B: torch.Tensor,
@@ -648,8 +488,7 @@ def invoke_fused_moe_kernel(
     A_scale: Optional[torch.Tensor],
     B_scale: Optional[torch.Tensor],
     B_zp: Optional[torch.Tensor],
-    topk_weights: torch.Tensor,
-    topk_ids: torch.Tensor,
+    topk_weights: Optional[torch.Tensor],
     sorted_token_ids: torch.Tensor,
     expert_ids: torch.Tensor,
     num_tokens_post_padded: torch.Tensor,
@@ -662,20 +501,21 @@ def invoke_fused_moe_kernel(
     use_int4_w4a16: bool,
     block_shape: Optional[List[int]] = None,
 ) -> None:
-    assert topk_weights.stride(1) == 1
+    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
-        if block_shape is None:
-            A, A_scale = scaled_fp8_quant(A, A_scale)
-        else:
-            assert len(block_shape) == 2
-            block_n, block_k = block_shape[0], block_shape[1]
-            A, A_scale = per_token_group_quant_fp8(A, block_k)
-            assert triton.cdiv(A.shape[-1], block_k) == A_scale.shape[-1]
-            assert triton.cdiv(B.shape[-2], block_n) == B_scale.shape[-2]
-            assert triton.cdiv(B.shape[-1], block_k) == B_scale.shape[-1]
+        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
@@ -683,6 +523,9 @@ def invoke_fused_moe_kernel(
         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.
@@ -703,6 +546,47 @@ def invoke_fused_moe_kernel(
         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,
@@ -716,7 +600,7 @@ def invoke_fused_moe_kernel(
             B.shape[1],
             A.shape[1],
             EM,
-            topk_ids.numel(),
+            num_tokens,
             A.stride(0),
             A.stride(1),
             B.stride(0),
@@ -740,8 +624,11 @@ def invoke_fused_moe_kernel(
             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,
@@ -753,9 +640,9 @@ def invoke_fused_moe_kernel(
             expert_ids,
             num_tokens_post_padded,
             B.shape[1],
-            A.shape[1],
+            B.shape[2],
             EM,
-            topk_ids.numel(),
+            num_tokens,
             A.stride(0),
             A.stride(1),
             B.stride(0),
@@ -775,6 +662,7 @@ def invoke_fused_moe_kernel(
             compute_type=compute_type,
             use_fp8_w8a8=use_fp8_w8a8,
             use_int8_w8a16=use_int8_w8a16,
+            BLOCK_SIZE_K=BLOCK_SIZE_K,
             **config,
         )
 
@@ -787,7 +675,7 @@ def get_config_file_name(
     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
 
 
@@ -835,6 +723,82 @@ def get_moe_configs(
     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,
@@ -856,6 +820,20 @@ def get_default_config(
             "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,
@@ -892,6 +870,8 @@ def try_get_optimal_moe_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)
@@ -908,12 +888,35 @@ def try_get_optimal_moe_config(
     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
@@ -926,17 +929,14 @@ def fused_topk(
         M, topk, dtype=torch.int32, device=hidden_states.device
     )
 
-    ops.topk_softmax(
-        topk_weights,
-        topk_ids,
-        token_expert_indicies,
-        gating_output.float(),  # TODO(woosuk): Optimize this.
-    )
-    del token_expert_indicies  # Not used. Will be used in the future.
+    gating_output_float = gating_output.float()  # TODO(woosuk): Optimize this.
 
-    if renormalize:
-        topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True)
+    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
 
 
@@ -951,7 +951,7 @@ def grouped_topk(
     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"
 
@@ -962,16 +962,19 @@ def grouped_topk(
     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)
-
-    num_token = scores.shape[0]
-    group_scores = (
-        scores.view(num_token, num_expert_group, -1).max(dim=-1).values
-    )  # [n, n_group]
+        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]
@@ -982,7 +985,7 @@ def grouped_topk(
         .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(), 0.0)  # [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]
@@ -1002,13 +1005,13 @@ def get_config_dtype_str(
     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_w8a16"
+        return "int4_w4a16"
     elif dtype == torch.float:
         # avoiding cases where kernel fails when float32 MoE
         # use fp16/bfloat16 configs
@@ -1022,9 +1025,13 @@ def inplace_fused_experts(
     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,
@@ -1040,9 +1047,13 @@ def inplace_fused_experts(
         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,
@@ -1053,15 +1064,51 @@ def inplace_fused_experts(
     )
 
 
+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,
@@ -1077,9 +1124,13 @@ def outplace_fused_experts(
         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,
@@ -1090,6 +1141,54 @@ def outplace_fused_experts(
     )
 
 
+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,
@@ -1097,9 +1196,13 @@ def fused_experts(
     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,
@@ -1107,44 +1210,28 @@ def fused_experts(
     a1_scale: Optional[torch.Tensor] = None,
     a2_scale: Optional[torch.Tensor] = None,
     block_shape: Optional[List[int]] = None,
-):
-    if inplace:
-        inplace_fused_experts(
-            hidden_states,
-            w1,
-            w2,
-            topk_weights,
-            topk_ids,
-            use_fp8_w8a8,
-            use_int8_w8a16,
-            use_int4_w4a16,
-            w1_scale,
-            w2_scale,
-            w1_zp,
-            w2_zp,
-            a1_scale,
-            a2_scale,
-            block_shape,
-        )
-        return hidden_states
-    else:
-        return outplace_fused_experts(
-            hidden_states,
-            w1,
-            w2,
-            topk_weights,
-            topk_ids,
-            use_fp8_w8a8,
-            use_int8_w8a16,
-            use_int4_w4a16,
-            w1_scale,
-            w2_scale,
-            w1_zp,
-            w2_zp,
-            a1_scale,
-            a2_scale,
-            block_shape,
-        )
+) -> 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(
@@ -1154,9 +1241,13 @@ def fused_experts_impl(
     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,
@@ -1173,12 +1264,16 @@ def fused_experts_impl(
 
     assert topk_weights.shape == topk_ids.shape, "topk shape 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 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
@@ -1194,27 +1289,26 @@ def fused_experts_impl(
         try_get_optimal_moe_config,
         w1.shape,
         w2.shape,
-        topk_ids.shape[1],
+        top_k_num,
         config_dtype,
         block_shape=block_shape,
     )
 
     config = get_config_func(M)
 
-    intermediate_cache1 = torch.empty(
-        (M, topk_ids.shape[1], N),
+    # 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 * topk_ids.shape[1], N // 2),
-        device=hidden_states.device,
-        dtype=hidden_states.dtype,
-    )
-    intermediate_cache3 = torch.empty(
-        (M, topk_ids.shape[1], w2.shape[1]),
-        device=hidden_states.device,
-        dtype=hidden_states.dtype,
+        (M * top_k_num, N // 2), device=hidden_states.device, dtype=hidden_states.dtype
     )
 
     if hidden_states.dtype == torch.bfloat16:
@@ -1248,31 +1342,42 @@ def fused_experts_impl(
             # 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]
+            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"], E
+            curr_topk_ids, config["BLOCK_SIZE_M"], global_num_experts, expert_map
         )
 
         invoke_fused_moe_kernel(
-            curr_hidden_states,
+            qcurr_hidden_states,
             w1,
             intermediate_cache1,
-            a1_scale,
+            a1q_scale,
             w1_scale,
             w1_zp,
             curr_topk_weights,
-            curr_topk_ids,
             sorted_token_ids,
             expert_ids,
             num_tokens_post_padded,
-            False,
-            topk_ids.shape[1],
+            apply_router_weight_on_input,
+            top_k_num,
             config,
             compute_type=compute_type,
             use_fp8_w8a8=use_fp8_w8a8,
@@ -1281,21 +1386,35 @@ def fused_experts_impl(
             block_shape=block_shape,
         )
 
-        ops.silu_and_mul(intermediate_cache2, intermediate_cache1.view(-1, N))
+        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(
-            intermediate_cache2,
+            qintermediate_cache2,
             w2,
             intermediate_cache3,
-            a2_scale,
+            a2q_scale,
             w2_scale,
             w2_zp,
             curr_topk_weights,
-            curr_topk_ids,
             sorted_token_ids,
             expert_ids,
             num_tokens_post_padded,
-            True,
+            not apply_router_weight_on_input,
             1,
             config,
             compute_type=compute_type,
@@ -1309,6 +1428,7 @@ def fused_experts_impl(
             intermediate_cache3.view(*intermediate_cache3.shape),
             out_hidden_states[begin_chunk_idx:end_chunk_idx],
         )
+
     return out_hidden_states
 
 
@@ -1320,6 +1440,8 @@ def fused_moe(
     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,
@@ -1327,6 +1449,8 @@ def fused_moe(
     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,
@@ -1349,6 +1473,8 @@ def fused_moe(
     - 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
@@ -1361,6 +1487,11 @@ def fused_moe(
     - 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
@@ -1375,8 +1506,6 @@ def fused_moe(
     Returns:
     - torch.Tensor: The output tensor after applying the MoE layer.
     """
-    # Check constraints.
-    assert gating_output.shape[1] == w1.shape[0], "Number of experts mismatch"
 
     if use_grouped_topk:
         assert num_expert_group is not None and topk_group is not None
@@ -1404,9 +1533,13 @@ def fused_moe(
         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,
diff --git a/torch-ext/moe/layers.py b/torch-ext/moe/layers.py
new file mode 100644
index 0000000000000000000000000000000000000000..0cf388ff61bd10cee921cd6f34bdc84a1b5025f6
--- /dev/null
+++ b/torch-ext/moe/layers.py
@@ -0,0 +1,52 @@
+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):
+    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)
+        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,
+        )
+
+        out += self.shared_expert(hidden_states)
+
+        return out, router_logits.t()
diff --git a/torch-ext/moe/moe_align_block_size.py b/torch-ext/moe/moe_align_block_size.py
new file mode 100644
index 0000000000000000000000000000000000000000..a698de5aaa123df5a2d663f252e1d5914e880d09
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/platforms.py b/torch-ext/moe/platforms.py
index 735fab87f2add390f7bf6408ebe31d1f5de6d02b..db83d0f2b0fadb15c14e8887ac27e4a07b6c4d62 100644
--- a/torch-ext/moe/platforms.py
+++ b/torch-ext/moe/platforms.py
@@ -1,6 +1,7 @@
+import random
 from abc import ABC, abstractmethod
 from functools import lru_cache
-from typing import NamedTuple
+from typing import NamedTuple, Optional
 
 import torch
 
@@ -31,9 +32,32 @@ class Platform(ABC):
     @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
@@ -46,6 +70,9 @@ class CudaPlatform(Platform):
     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
 
@@ -62,6 +89,9 @@ class RocmPlatform(Platform):
     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
 
diff --git a/torch-ext/moe/utils/custom_ops.by b/torch-ext/moe/utils/custom_ops.by
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/torch-ext/moe/utils/custom_ops.py b/torch-ext/moe/utils/custom_ops.py
new file mode 100644
index 0000000000000000000000000000000000000000..26ef7c23d1bea7a162d81746ac041baefb472f46
--- /dev/null
+++ b/torch-ext/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/torch-ext/moe/utils/marlin_utils.py b/torch-ext/moe/utils/marlin_utils.py
index 5037f774b8a8b7e88d822efacbb3b4ea5b95d356..f27b71af67a6f5f18c8dba271926e3d40d518879 100644
--- a/torch-ext/moe/utils/marlin_utils.py
+++ b/torch-ext/moe/utils/marlin_utils.py
@@ -2,8 +2,8 @@ from typing import List, Optional, Tuple
 
 import numpy
 import torch
-from moe.scalar_type import ScalarType, scalar_types
 
+from ..scalar_type import ScalarType, scalar_types
 from .quant_utils import pack_cols, unpack_cols
 
 GPTQ_MARLIN_TILE = 16
diff --git a/torch-ext/moe/utils/marlin_utils_test.py b/torch-ext/moe/utils/marlin_utils_test.py
index 83faac032ca93b3564c620c5b4b1ef63c74aaddf..b03555f5bffa8d48187d37e53f17e6e3138472cc 100644
--- a/torch-ext/moe/utils/marlin_utils_test.py
+++ b/torch-ext/moe/utils/marlin_utils_test.py
@@ -4,8 +4,8 @@ from typing import List, Optional
 
 import numpy as np
 import torch
-from moe.scalar_type import ScalarType
 
+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,
diff --git a/torch-ext/moe/utils/quant_utils.py b/torch-ext/moe/utils/quant_utils.py
index 5819ab753e57655185572ce1e49c24e6268171b4..65c90821773b60cca2b0f9102ca7ad4e3cd009d5 100644
--- a/torch-ext/moe/utils/quant_utils.py
+++ b/torch-ext/moe/utils/quant_utils.py
@@ -4,7 +4,8 @@ from typing import List, Optional
 
 import numpy
 import torch
-from moe.scalar_type import ScalarType, scalar_types
+
+from ..scalar_type import ScalarType, scalar_types
 
 SUPPORTED_GPTQ_QUANT_TYPES = [scalar_types.uint4b8, scalar_types.uint8b128]
 SUPPORTED_GROUP_SIZES = [-1, 32, 64, 128]
diff --git a/torch-ext/registration.h b/torch-ext/registration.h
new file mode 100644
index 0000000000000000000000000000000000000000..19a82cca9a230115263ce506f0476096ebd3a226
--- /dev/null
+++ b/torch-ext/registration.h
@@ -0,0 +1,30 @@
+// Registration macros from vLLM:
+// https://github.com/vllm-project/vllm/blob/main/csrc/core/registration.h
+
+#pragma once
+
+#include <Python.h>
+
+#define _CONCAT(A, B) A##B
+#define CONCAT(A, B) _CONCAT(A, B)
+
+#define _STRINGIFY(A) #A
+#define STRINGIFY(A) _STRINGIFY(A)
+
+// A version of the TORCH_LIBRARY macro that expands the NAME, i.e. so NAME
+// could be a macro instead of a literal token.
+#define TORCH_LIBRARY_EXPAND(NAME, MODULE) TORCH_LIBRARY(NAME, MODULE)
+
+// A version of the TORCH_LIBRARY_IMPL macro that expands the NAME, i.e. so NAME
+// could be a macro instead of a literal token.
+#define TORCH_LIBRARY_IMPL_EXPAND(NAME, DEVICE, MODULE) \
+  TORCH_LIBRARY_IMPL(NAME, DEVICE, MODULE)
+
+// REGISTER_EXTENSION allows the shared library to be loaded and initialized
+// via python's import statement.
+#define REGISTER_EXTENSION(NAME)                                               \
+  PyMODINIT_FUNC CONCAT(PyInit_, NAME)() {                                     \
+    static struct PyModuleDef module = {PyModuleDef_HEAD_INIT,                 \
+                                        STRINGIFY(NAME), nullptr, 0, nullptr}; \
+    return PyModule_Create(&module);                                           \
+  }
diff --git a/torch-ext/torch_binding.cpp b/torch-ext/torch_binding.cpp
index 2303c217ef6b159c08cd9b974bced721d9badab3..daff107eed6443511b02fa9fb9b41fb083d8d3a3 100644
--- a/torch-ext/torch_binding.cpp
+++ b/torch-ext/torch_binding.cpp
@@ -55,6 +55,16 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
            &dynamic_per_token_scaled_fp8_quant);
 
 #ifndef USE_ROCM
+  ops.def(
+      "moe_wna16_gemm(Tensor input, Tensor! output, Tensor b_qweight, "
+      "Tensor b_scales, Tensor? b_qzeros, "
+      "Tensor? topk_weights, Tensor sorted_token_ids, "
+      "Tensor expert_ids, Tensor num_tokens_post_pad, "
+      "int top_k, int BLOCK_SIZE_M, int BLOCK_SIZE_N, int BLOCK_SIZE_K, "
+      "int bit) -> Tensor");
+
+  ops.impl("moe_wna16_gemm", torch::kCUDA, &moe_wna16_gemm);
+
   ops.def("marlin_gemm_moe(Tensor! a, Tensor! b_q_weights, Tensor! sorted_ids, "
           "Tensor! topk_weights, Tensor! topk_ids, Tensor! b_scales, Tensor! "
           "b_zeros, Tensor! g_idx, Tensor! perm, Tensor! workspace, "
diff --git a/torch-ext/torch_binding.h b/torch-ext/torch_binding.h
index 2e59e6d043abbe9d018c22840095a80d39e4acb6..aec3cf985290c36a8a7cf726878c38ed648dd24a 100644
--- a/torch-ext/torch_binding.h
+++ b/torch-ext/torch_binding.h
@@ -34,6 +34,16 @@ void dynamic_per_token_scaled_fp8_quant(
     std::optional<torch::Tensor> const &scale_ub);
 
 #ifndef USE_ROCM
+torch::Tensor moe_wna16_gemm(torch::Tensor input, torch::Tensor output,
+                             torch::Tensor b_qweight, torch::Tensor b_scales,
+                             std::optional<torch::Tensor> b_qzeros,
+                             std::optional<torch::Tensor> topk_weights,
+                             torch::Tensor sorted_token_ids,
+                             torch::Tensor expert_ids,
+                             torch::Tensor num_tokens_post_pad, int64_t top_k,
+                             int64_t BLOCK_SIZE_M, int64_t BLOCK_SIZE_N,
+                             int64_t BLOCK_SIZE_K, int64_t bit);
+
 torch::Tensor marlin_gemm_moe(
     const torch::Tensor &a, const torch::Tensor &b_q_weights,
     const torch::Tensor &sorted_ids, const torch::Tensor &topk_weights,