biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Hardware][NVIDIA] FP4 MoE kernel optimization (#19110)

Browse Source 
Signed-off-by: Chiyue Wei <chiyuew@nvidia.com>
Co-authored-by: Chiyue Wei <chiyuew@nvidia.com>
This commit is contained in:
Chiyue Wei
2025-06-05 09:48:26 -07:00
committed by GitHub
parent ec89524f50
commit 61059bee40
12 changed files with 165 additions and 38 deletions
Download Patch File Download Diff File Expand all files Collapse all files

36
csrc/quantization/cutlass_w8a8/moe/moe_data.cu
View File

@@ -45,6 +45,23 @@ __global__ void compute_expert_offsets(
}
}
__global__ void compute_expert_blockscale_offsets(
const int32_t* __restrict__ problem_sizes1, int32_t* expert_offsets,
int32_t* blockscale_offsets, int32_t* atomic_buffer,
const int num_experts) {
int32_t tot_offset = 0;
int32_t tot_offset_round = 0;
expert_offsets[0] = 0;
blockscale_offsets[0] = 0;
for (int i = 0; i < num_experts; ++i) {
atomic_buffer[i] = tot_offset;
tot_offset += problem_sizes1[i * 3];
expert_offsets[i + 1] = tot_offset;
tot_offset_round += (problem_sizes1[i * 3] + (128 - 1)) / 128 * 128;
blockscale_offsets[i + 1] = tot_offset_round;
}
}
__global__ void compute_arg_sorts(const int* __restrict__ topk_ids,
const int32_t* __restrict__ expert_offsets,
int32_t* input_permutation,
@@ -77,7 +94,8 @@ void get_cutlass_moe_mm_data_caller(
const torch::Tensor& topk_ids, torch::Tensor& expert_offsets,
torch::Tensor& problem_sizes1, torch::Tensor& problem_sizes2,
torch::Tensor& input_permutation, torch::Tensor& output_permutation,
const int64_t num_experts, const int64_t n, const int64_t k) {
const int64_t num_experts, const int64_t n, const int64_t k,
const std::optional<torch::Tensor>& blockscale_offsets) {
auto stream = at::cuda::getCurrentCUDAStream(topk_ids.device().index());
auto options_int32 =
torch::TensorOptions().dtype(torch::kInt32).device(topk_ids.device());
@@ -89,10 +107,18 @@ void get_cutlass_moe_mm_data_caller(
static_cast<int32_t*>(problem_sizes1.data_ptr()),
static_cast<int32_t*>(problem_sizes2.data_ptr()),
static_cast<int32_t*>(atomic_buffer.data_ptr()), topk_ids.numel(), n, k);
compute_expert_offsets<<<1, 1, 0, stream>>>(
static_cast<const int32_t*>(problem_sizes1.data_ptr()),
static_cast<int32_t*>(expert_offsets.data_ptr()),
static_cast<int32_t*>(atomic_buffer.data_ptr()), num_experts);
if (blockscale_offsets.has_value()) {
compute_expert_blockscale_offsets<<<1, 1, 0, stream>>>(
static_cast<const int32_t*>(problem_sizes1.data_ptr()),
static_cast<int32_t*>(expert_offsets.data_ptr()),
static_cast<int32_t*>(blockscale_offsets.value().data_ptr()),
static_cast<int32_t*>(atomic_buffer.data_ptr()), num_experts);
} else {
compute_expert_offsets<<<1, 1, 0, stream>>>(
static_cast<const int32_t*>(problem_sizes1.data_ptr()),
static_cast<int32_t*>(expert_offsets.data_ptr()),
static_cast<int32_t*>(atomic_buffer.data_ptr()), num_experts);
}
compute_arg_sorts<<<num_experts, num_threads, 0, stream>>>(
static_cast<const int32_t*>(topk_ids.data_ptr()),
static_cast<const int32_t*>(expert_offsets.data_ptr()),