[Kernel] Integrate CUTLASS MoE kernel with PPLX (#18762)
Signed-off-by: ElizaWszola <ewszola@redhat.com> Signed-off-by: Tyler Michael Smith <tyler@neuralmagic.com> Co-authored-by: Tyler Michael Smith <tyler@neuralmagic.com>
This commit is contained in:
co-authored by
Tyler Michael Smith
parent
6e0cd10f72
commit
84166fee97
+10
-1
@@ -236,7 +236,8 @@ void cutlass_moe_mm(
|
||||
torch::Tensor const& b_tensors, torch::Tensor const& a_scales,
|
||||
torch::Tensor const& b_scales, torch::Tensor const& expert_offsets,
|
||||
torch::Tensor const& problem_sizes, torch::Tensor const& a_strides,
|
||||
torch::Tensor const& b_strides, torch::Tensor const& c_strides);
|
||||
torch::Tensor const& b_strides, torch::Tensor const& c_strides,
|
||||
bool per_act_token, bool per_out_ch);
|
||||
|
||||
void cutlass_fp4_group_mm(
|
||||
torch::Tensor& output, const torch::Tensor& a, const torch::Tensor& b,
|
||||
@@ -251,6 +252,14 @@ void get_cutlass_moe_mm_data(
|
||||
const int64_t num_experts, const int64_t n, const int64_t k,
|
||||
const std::optional<torch::Tensor>& blockscale_offsets);
|
||||
|
||||
void get_cutlass_pplx_moe_mm_data(torch::Tensor& expert_offsets,
|
||||
torch::Tensor& problem_sizes1,
|
||||
torch::Tensor& problem_sizes2,
|
||||
const torch::Tensor& expert_num_tokens,
|
||||
const int64_t num_local_experts,
|
||||
const int64_t padded_m, const int64_t n,
|
||||
const int64_t k);
|
||||
|
||||
void cutlass_scaled_mm_azp(torch::Tensor& out, torch::Tensor const& a,
|
||||
torch::Tensor const& b,
|
||||
torch::Tensor const& a_scales,
|
||||
|
||||
@@ -84,7 +84,8 @@ void run_cutlass_moe_mm_sm90(
|
||||
torch::Tensor const& b_tensors, torch::Tensor const& a_scales,
|
||||
torch::Tensor const& b_scales, torch::Tensor const& expert_offsets,
|
||||
torch::Tensor const& problem_sizes, torch::Tensor const& a_strides,
|
||||
torch::Tensor const& b_strides, torch::Tensor const& c_strides) {
|
||||
torch::Tensor const& b_strides, torch::Tensor const& c_strides,
|
||||
bool per_act_token, bool per_out_ch) {
|
||||
TORCH_CHECK(a_tensors.size(0) > 0, "No input A tensors provided.");
|
||||
TORCH_CHECK(b_tensors.size(0) > 0, "No input B tensors provided.");
|
||||
TORCH_CHECK(out_tensors.size(0) > 0, "No output tensors provided.");
|
||||
@@ -113,19 +114,23 @@ void run_cutlass_moe_mm_sm90(
|
||||
if (n >= 8192) {
|
||||
cutlass_group_gemm_caller<Cutlass3xGemmN8192>(
|
||||
out_tensors, a_tensors, b_tensors, a_scales, b_scales, expert_offsets,
|
||||
problem_sizes, a_strides, b_strides, c_strides);
|
||||
problem_sizes, a_strides, b_strides, c_strides, per_act_token,
|
||||
per_out_ch);
|
||||
} else if (k >= 8192) {
|
||||
cutlass_group_gemm_caller<Cutlass3xGemmK8192>(
|
||||
out_tensors, a_tensors, b_tensors, a_scales, b_scales, expert_offsets,
|
||||
problem_sizes, a_strides, b_strides, c_strides);
|
||||
problem_sizes, a_strides, b_strides, c_strides, per_act_token,
|
||||
per_out_ch);
|
||||
} else if (m <= 16) {
|
||||
cutlass_group_gemm_caller<Cutlass3xGemmM16>(
|
||||
out_tensors, a_tensors, b_tensors, a_scales, b_scales, expert_offsets,
|
||||
problem_sizes, a_strides, b_strides, c_strides);
|
||||
problem_sizes, a_strides, b_strides, c_strides, per_act_token,
|
||||
per_out_ch);
|
||||
} else {
|
||||
cutlass_group_gemm_caller<Cutlass3xGemmDefault>(
|
||||
out_tensors, a_tensors, b_tensors, a_scales, b_scales, expert_offsets,
|
||||
problem_sizes, a_strides, b_strides, c_strides);
|
||||
problem_sizes, a_strides, b_strides, c_strides, per_act_token,
|
||||
per_out_ch);
|
||||
}
|
||||
}
|
||||
|
||||
@@ -134,15 +139,18 @@ void dispatch_moe_mm_sm90(
|
||||
torch::Tensor const& b_tensors, torch::Tensor const& a_scales,
|
||||
torch::Tensor const& b_scales, torch::Tensor const& expert_offsets,
|
||||
torch::Tensor const& problem_sizes, torch::Tensor const& a_strides,
|
||||
torch::Tensor const& b_strides, torch::Tensor const& c_strides) {
|
||||
torch::Tensor const& b_strides, torch::Tensor const& c_strides,
|
||||
bool per_act_token, bool per_out_ch) {
|
||||
if (out_tensors.dtype() == torch::kBFloat16) {
|
||||
run_cutlass_moe_mm_sm90<cutlass::float_e4m3_t, cutlass::bfloat16_t>(
|
||||
out_tensors, a_tensors, b_tensors, a_scales, b_scales, expert_offsets,
|
||||
problem_sizes, a_strides, b_strides, c_strides);
|
||||
problem_sizes, a_strides, b_strides, c_strides, per_act_token,
|
||||
per_out_ch);
|
||||
} else {
|
||||
run_cutlass_moe_mm_sm90<cutlass::float_e4m3_t, cutlass::half_t>(
|
||||
out_tensors, a_tensors, b_tensors, a_scales, b_scales, expert_offsets,
|
||||
problem_sizes, a_strides, b_strides, c_strides);
|
||||
problem_sizes, a_strides, b_strides, c_strides, per_act_token,
|
||||
per_out_ch);
|
||||
}
|
||||
}
|
||||
|
||||
@@ -153,8 +161,9 @@ void cutlass_moe_mm_sm90(
|
||||
torch::Tensor const& b_tensors, torch::Tensor const& a_scales,
|
||||
torch::Tensor const& b_scales, torch::Tensor const& expert_offsets,
|
||||
torch::Tensor const& problem_sizes, torch::Tensor const& a_strides,
|
||||
torch::Tensor const& b_strides, torch::Tensor const& c_strides) {
|
||||
torch::Tensor const& b_strides, torch::Tensor const& c_strides,
|
||||
bool per_act_token, bool per_out_ch) {
|
||||
dispatch_moe_mm_sm90(out_tensors, a_tensors, b_tensors, a_scales, b_scales,
|
||||
expert_offsets, problem_sizes, a_strides, b_strides,
|
||||
c_strides);
|
||||
c_strides, per_act_token, per_out_ch);
|
||||
}
|
||||
|
||||
@@ -76,7 +76,8 @@ void cutlass_group_gemm_caller(
|
||||
torch::Tensor const& b_tensors, torch::Tensor const& a_scales,
|
||||
torch::Tensor const& b_scales, torch::Tensor const& expert_offsets,
|
||||
torch::Tensor const& problem_sizes, torch::Tensor const& a_strides,
|
||||
torch::Tensor const& b_strides, torch::Tensor const& c_strides) {
|
||||
torch::Tensor const& b_strides, torch::Tensor const& c_strides,
|
||||
bool per_act_token, bool per_out_ch) {
|
||||
using ElementAB = typename Gemm::ElementAB;
|
||||
using ElementD = typename Gemm::ElementD;
|
||||
|
||||
@@ -84,9 +85,6 @@ void cutlass_group_gemm_caller(
|
||||
int k_size = a_tensors.size(1);
|
||||
int n_size = out_tensors.size(1);
|
||||
|
||||
bool per_act_token = a_scales.numel() != 1;
|
||||
bool per_out_ch = b_scales.numel() != num_experts;
|
||||
|
||||
auto stream = at::cuda::getCurrentCUDAStream(a_tensors.device().index());
|
||||
|
||||
auto options_int =
|
||||
|
||||
@@ -7,7 +7,7 @@
|
||||
|
||||
constexpr uint64_t THREADS_PER_EXPERT = 512;
|
||||
|
||||
__global__ void compute_problem_sizes(const int* __restrict__ topk_ids,
|
||||
__global__ void compute_problem_sizes(const uint32_t* __restrict__ topk_ids,
|
||||
int32_t* problem_sizes1,
|
||||
int32_t* problem_sizes2,
|
||||
int32_t* atomic_buffer,
|
||||
@@ -62,7 +62,7 @@ __global__ void compute_expert_blockscale_offsets(
|
||||
}
|
||||
}
|
||||
|
||||
__global__ void compute_arg_sorts(const int* __restrict__ topk_ids,
|
||||
__global__ void compute_arg_sorts(const uint32_t* __restrict__ topk_ids,
|
||||
const int32_t* __restrict__ expert_offsets,
|
||||
int32_t* input_permutation,
|
||||
int32_t* output_permutation,
|
||||
@@ -103,7 +103,7 @@ void get_cutlass_moe_mm_data_caller(
|
||||
|
||||
int num_threads = min(THREADS_PER_EXPERT, topk_ids.numel());
|
||||
compute_problem_sizes<<<num_experts, num_threads, 0, stream>>>(
|
||||
static_cast<const int32_t*>(topk_ids.data_ptr()),
|
||||
static_cast<const uint32_t*>(topk_ids.data_ptr()),
|
||||
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);
|
||||
@@ -120,10 +120,44 @@ void get_cutlass_moe_mm_data_caller(
|
||||
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 uint32_t*>(topk_ids.data_ptr()),
|
||||
static_cast<const int32_t*>(expert_offsets.data_ptr()),
|
||||
static_cast<int32_t*>(input_permutation.data_ptr()),
|
||||
static_cast<int32_t*>(output_permutation.data_ptr()),
|
||||
static_cast<int32_t*>(atomic_buffer.data_ptr()), topk_ids.numel(),
|
||||
topk_ids.size(1));
|
||||
}
|
||||
|
||||
__global__ void compute_pplx_data(int32_t* expert_offsets,
|
||||
int32_t* problem_sizes1,
|
||||
int32_t* problem_sizes2,
|
||||
const int32_t* __restrict__ expert_num_tokens,
|
||||
const int padded_m, const int n,
|
||||
const int k) {
|
||||
int expert_idx = threadIdx.x;
|
||||
|
||||
expert_offsets[expert_idx] = expert_idx * padded_m;
|
||||
problem_sizes1[expert_idx * 3] = expert_num_tokens[expert_idx];
|
||||
problem_sizes1[expert_idx * 3 + 1] = 2 * n;
|
||||
problem_sizes1[expert_idx * 3 + 2] = k;
|
||||
problem_sizes2[expert_idx * 3] = expert_num_tokens[expert_idx];
|
||||
problem_sizes2[expert_idx * 3 + 1] = k;
|
||||
problem_sizes2[expert_idx * 3 + 2] = n;
|
||||
}
|
||||
|
||||
void get_cutlass_pplx_moe_mm_data_caller(torch::Tensor& expert_offsets,
|
||||
torch::Tensor& problem_sizes1,
|
||||
torch::Tensor& problem_sizes2,
|
||||
const torch::Tensor& expert_num_tokens,
|
||||
const int64_t num_local_experts,
|
||||
const int64_t padded_m,
|
||||
const int64_t n, const int64_t k) {
|
||||
auto stream = at::cuda::getCurrentCUDAStream(expert_offsets.device().index());
|
||||
|
||||
compute_pplx_data<<<1, num_local_experts, 0, stream>>>(
|
||||
static_cast<int32_t*>(expert_offsets.data_ptr()),
|
||||
static_cast<int32_t*>(problem_sizes1.data_ptr()),
|
||||
static_cast<int32_t*>(problem_sizes2.data_ptr()),
|
||||
static_cast<const int32_t*>(expert_num_tokens.data_ptr()), padded_m, n,
|
||||
k);
|
||||
}
|
||||
|
||||
@@ -36,7 +36,8 @@ void cutlass_moe_mm_sm90(
|
||||
torch::Tensor const& b_tensors, torch::Tensor const& a_scales,
|
||||
torch::Tensor const& b_scales, torch::Tensor const& expert_offsets,
|
||||
torch::Tensor const& problem_sizes, torch::Tensor const& a_strides,
|
||||
torch::Tensor const& b_strides, torch::Tensor const& c_strides);
|
||||
torch::Tensor const& b_strides, torch::Tensor const& c_strides,
|
||||
bool per_act_token, bool per_out_ch);
|
||||
|
||||
#endif
|
||||
|
||||
@@ -56,6 +57,14 @@ void get_cutlass_moe_mm_data_caller(
|
||||
torch::Tensor& input_permutation, torch::Tensor& output_permutation,
|
||||
const int64_t num_experts, const int64_t n, const int64_t k,
|
||||
const std::optional<torch::Tensor>& blockscale_offsets);
|
||||
|
||||
void get_cutlass_pplx_moe_mm_data_caller(torch::Tensor& expert_offsets,
|
||||
torch::Tensor& problem_sizes1,
|
||||
torch::Tensor& problem_sizes2,
|
||||
const torch::Tensor& expert_num_tokens,
|
||||
const int64_t num_local_experts,
|
||||
const int64_t padded_m,
|
||||
const int64_t n, const int64_t k);
|
||||
#endif
|
||||
|
||||
void cutlass_scaled_mm_azp_sm75(torch::Tensor& c, torch::Tensor const& a,
|
||||
@@ -207,12 +216,13 @@ void cutlass_moe_mm(
|
||||
torch::Tensor const& b_tensors, torch::Tensor const& a_scales,
|
||||
torch::Tensor const& b_scales, torch::Tensor const& expert_offsets,
|
||||
torch::Tensor const& problem_sizes, torch::Tensor const& a_strides,
|
||||
torch::Tensor const& b_strides, torch::Tensor const& c_strides) {
|
||||
torch::Tensor const& b_strides, torch::Tensor const& c_strides,
|
||||
bool per_act_token, bool per_out_ch) {
|
||||
int32_t version_num = get_sm_version_num();
|
||||
#if defined ENABLE_CUTLASS_MOE_SM90 && ENABLE_CUTLASS_MOE_SM90
|
||||
cutlass_moe_mm_sm90(out_tensors, a_tensors, b_tensors, a_scales, b_scales,
|
||||
expert_offsets, problem_sizes, a_strides, b_strides,
|
||||
c_strides);
|
||||
c_strides, per_act_token, per_out_ch);
|
||||
return;
|
||||
#endif
|
||||
TORCH_CHECK_NOT_IMPLEMENTED(
|
||||
@@ -245,6 +255,29 @@ void get_cutlass_moe_mm_data(
|
||||
version_num, ". Required capability: 90");
|
||||
}
|
||||
|
||||
void get_cutlass_pplx_moe_mm_data(torch::Tensor& expert_offsets,
|
||||
torch::Tensor& problem_sizes1,
|
||||
torch::Tensor& problem_sizes2,
|
||||
const torch::Tensor& expert_num_tokens,
|
||||
const int64_t num_local_experts,
|
||||
const int64_t padded_m, const int64_t n,
|
||||
const int64_t k) {
|
||||
// This function currently gets compiled only if we have a valid cutlass moe
|
||||
// mm to run it for.
|
||||
int32_t version_num = get_sm_version_num();
|
||||
#if defined ENABLE_CUTLASS_MOE_SM90 && ENABLE_CUTLASS_MOE_SM90
|
||||
get_cutlass_pplx_moe_mm_data_caller(expert_offsets, problem_sizes1,
|
||||
problem_sizes2, expert_num_tokens,
|
||||
num_local_experts, padded_m, n, k);
|
||||
return;
|
||||
#endif
|
||||
TORCH_CHECK_NOT_IMPLEMENTED(
|
||||
false,
|
||||
"No compiled get_cutlass_pplx_moe_mm_data: no cutlass_scaled_mm kernel "
|
||||
"for CUDA device capability: ",
|
||||
version_num, ". Required capability: 90");
|
||||
}
|
||||
|
||||
void cutlass_scaled_mm_azp(torch::Tensor& c, torch::Tensor const& a,
|
||||
torch::Tensor const& b,
|
||||
torch::Tensor const& a_scales,
|
||||
|
||||
+18
-1
@@ -435,7 +435,8 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
|
||||
"cutlass_moe_mm(Tensor! out_tensors, Tensor a_tensors, Tensor b_tensors, "
|
||||
" Tensor a_scales, Tensor b_scales, Tensor expert_offsets, "
|
||||
" Tensor problem_sizes, Tensor a_strides, "
|
||||
" Tensor b_strides, Tensor c_strides) -> ()",
|
||||
" Tensor b_strides, Tensor c_strides, bool per_act_token, "
|
||||
" bool per_out_ch) -> ()",
|
||||
{stride_tag});
|
||||
ops.impl("cutlass_moe_mm", torch::kCUDA, &cutlass_moe_mm);
|
||||
|
||||
@@ -454,6 +455,22 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
|
||||
{stride_tag});
|
||||
ops.impl("get_cutlass_moe_mm_data", torch::kCUDA, &get_cutlass_moe_mm_data);
|
||||
|
||||
// A function that computes data required to run fused MoE with w8a8 grouped
|
||||
// GEMM and PPLX. It takes expert_num_tokens and non_zero_expert_idxs
|
||||
// as an input, and computes expert_offsets (token start indices of each
|
||||
// expert). In addition to this, it computes problem sizes for each expert's
|
||||
// multiplication used by the two mms called from fused MoE operation.
|
||||
ops.def(
|
||||
"get_cutlass_pplx_moe_mm_data(Tensor! expert_offsets, "
|
||||
" Tensor! problem_sizes1, "
|
||||
" Tensor! problem_sizes2, "
|
||||
" Tensor expert_num_tokens, "
|
||||
" int num_local_experts, int padded_m, "
|
||||
" int n, int k) -> ()",
|
||||
{stride_tag});
|
||||
ops.impl("get_cutlass_pplx_moe_mm_data", torch::kCUDA,
|
||||
&get_cutlass_pplx_moe_mm_data);
|
||||
|
||||
// Check if cutlass scaled_mm supports block quantization (used by DeepSeekV3)
|
||||
ops.def(
|
||||
"cutlass_scaled_mm_supports_block_fp8(int cuda_device_capability) -> "
|
||||
|
||||
Reference in New Issue
Block a user