Skip to content

GitLab

Unverified Commit ab1a6a43 authored by mikaylagawarecki's avatar mikaylagawarecki Committed by GitHub
Browse files

[3/n] Migrate cutlass/scaled_mm_entry.cu torch stable ABI (#37221) 


Signed-off-by: default avatarMikayla Gawarecki <mikaylagawarecki@gmail.com>
parent b5e60825
Hide whitespace changes
Inline Side-by-side
Showing with 6 additions and 185 deletions
csrc/libtorch_stable/torch_utils.h
View file @ ab1a6a43
#pragma once
#include <torch/csrc/inductor/aoti_torch/c/shim.h>
#include <torch/csrc/stable/accelerator.h>
#include <torch/csrc/stable/tensor.h>
#include <torch/headeronly/util/shim_utils.h>
#include <cuda_runtime.h>
// Stable ABI equivalent of TORCH_CHECK_NOT_IMPLEMENTED.
#define STD_TORCH_CHECK_NOT_IMPLEMENTED(cond, ...) \
STD_TORCH_CHECK(cond, "NotImplementedError: ", __VA_ARGS__)
// Utility to get the current CUDA stream for a given device using stable APIs.
// Returns a cudaStream_t for use in kernel launches.
inline cudaStream_t get_current_cuda_stream(int32_t device_index = -1) {
......
csrc/ops.h
View file @ ab1a6a43
......@@ -228,63 +228,18 @@ int64_t ggml_moe_get_block_size(int64_t type);
#ifndef USE_ROCM
bool cutlass_scaled_mm_supports_fp4(int64_t cuda_device_capability);
bool cutlass_scaled_mm_supports_fp8(int64_t cuda_device_capability);
bool cutlass_scaled_mm_supports_block_fp8(int64_t cuda_device_capability);
bool cutlass_group_gemm_supported(int64_t cuda_device_capability);
void cutlass_scaled_fp4_mm(torch::Tensor& D, torch::Tensor const& A,
torch::Tensor const& B, torch::Tensor const& A_sf,
torch::Tensor const& B_sf,
torch::Tensor const& alpha);
void cutlass_scaled_mm(torch::Tensor& out, torch::Tensor const& a,
torch::Tensor const& b, torch::Tensor const& a_scales,
torch::Tensor const& b_scales,
std::optional<torch::Tensor> const& bias);
void cutlass_moe_mm(
torch::Tensor& out_tensors, torch::Tensor const& a_tensors,
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,
bool per_act_token, bool per_out_ch);
void cutlass_fp4_group_mm(
torch::Tensor& output, const torch::Tensor& a, const torch::Tensor& b,
const torch::Tensor& a_blockscale, const torch::Tensor& b_blockscales,
const torch::Tensor& alphas, const torch::Tensor& problem_sizes,
const torch::Tensor& expert_offsets, const torch::Tensor& sf_offsets);
void get_cutlass_moe_mm_data(
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 std::optional<torch::Tensor>& blockscale_offsets,
const bool is_gated);
void get_cutlass_moe_mm_problem_sizes_from_expert_offsets(
const torch::Tensor& expert_first_token_offset,
torch::Tensor& problem_sizes1, torch::Tensor& problem_sizes2,
const int64_t n, const int64_t k, const bool swap_ab);
void get_cutlass_batched_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,
torch::Tensor const& b_scales,
torch::Tensor const& azp_adj,
std::optional<torch::Tensor> const& azp,
std::optional<torch::Tensor> const& bias);
std::tuple<torch::Tensor, torch::Tensor> scaled_fp4_quant_func(
torch::Tensor const& input, torch::Tensor const& input_scale,
bool is_sf_swizzled_layout);
......
csrc/quantization/w8a8/cutlass/c3x/scaled_mm_kernels.hpp deleted 100644 → 0
View file @ b5e60825
#pragma once
#include <torch/all.h>
namespace vllm {
void cutlass_scaled_mm_sm90_fp8(torch::Tensor& out, torch::Tensor const& a,
torch::Tensor const& b,
torch::Tensor const& a_scales,
torch::Tensor const& b_scales,
std::optional<torch::Tensor> const& bias);
void cutlass_scaled_mm_sm90_int8(torch::Tensor& out, torch::Tensor const& a,
torch::Tensor const& b,
torch::Tensor const& a_scales,
torch::Tensor const& b_scales,
std::optional<torch::Tensor> const& bias);
void cutlass_scaled_mm_azp_sm90_int8(torch::Tensor& out, torch::Tensor const& a,
torch::Tensor const& b,
torch::Tensor const& a_scales,
torch::Tensor const& b_scales,
torch::Tensor const& azp_adj,
std::optional<torch::Tensor> const& azp,
std::optional<torch::Tensor> const& bias);
void cutlass_scaled_mm_blockwise_sm90_fp8(torch::Tensor& out,
torch::Tensor const& a,
torch::Tensor const& b,
torch::Tensor const& a_scales,
torch::Tensor const& b_scales);
void cutlass_scaled_mm_sm100_fp8(torch::Tensor& out, torch::Tensor const& a,
torch::Tensor const& b,
torch::Tensor const& a_scales,
torch::Tensor const& b_scales,
std::optional<torch::Tensor> const& bias);
void cutlass_scaled_mm_sm120_fp8(torch::Tensor& out, torch::Tensor const& a,
torch::Tensor const& b,
torch::Tensor const& a_scales,
torch::Tensor const& b_scales,
std::optional<torch::Tensor> const& bias);
void cutlass_scaled_mm_blockwise_sm100_fp8(torch::Tensor& out,
torch::Tensor const& a,
torch::Tensor const& b,
torch::Tensor const& a_scales,
torch::Tensor const& b_scales);
void cutlass_scaled_mm_blockwise_sm120_fp8(torch::Tensor& out,
torch::Tensor const& a,
torch::Tensor const& b,
torch::Tensor const& a_scales,
torch::Tensor const& b_scales);
} // namespace vllm
csrc/torch_bindings.cpp
View file @ ab1a6a43
......@@ -439,90 +439,6 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
" -> ()");
// conditionally compiled so impl registration is in source file
// CUTLASS w8a8 GEMM, supporting symmetric per-tensor or per-row/column
// quantization, as well as bias
ops.def(
"cutlass_scaled_mm(Tensor! out, Tensor a,"
" Tensor b, Tensor a_scales,"
" Tensor b_scales, Tensor? bias) -> ()");
ops.impl("cutlass_scaled_mm", torch::kCUDA, &cutlass_scaled_mm);
// CUTLASS w8a8 GEMM, supporting asymmetric per-tensor or per-row/column
// quantization.
ops.def(
"cutlass_scaled_mm_azp(Tensor! out, Tensor a,"
" Tensor b, Tensor a_scales,"
" Tensor b_scales, Tensor azp_adj,"
" Tensor? azp, Tensor? bias) -> ()");
ops.impl("cutlass_scaled_mm_azp", torch::kCUDA, &cutlass_scaled_mm_azp);
// Check if cutlass scaled_mm is supported for CUDA devices of the given
// capability
ops.def("cutlass_scaled_mm_supports_fp8(int cuda_device_capability) -> bool");
ops.impl("cutlass_scaled_mm_supports_fp8", &cutlass_scaled_mm_supports_fp8);
// Check if cutlass grouped gemm is supported for CUDA devices of the given
// capability
ops.def("cutlass_group_gemm_supported(int cuda_device_capability) -> bool");
ops.impl("cutlass_group_gemm_supported", &cutlass_group_gemm_supported);
// CUTLASS w8a8 grouped GEMM
ops.def(
"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, bool per_act_token, "
" bool per_out_ch) -> ()");
ops.impl("cutlass_moe_mm", torch::kCUDA, &cutlass_moe_mm);
// A function that computes data required to run fused MoE with w8a8 grouped
// GEMM. It takes topk_ids 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, and arrays with permutations required to shuffle
// and de-shuffle the input/output of the fused operation.
ops.def(
"get_cutlass_moe_mm_data(Tensor topk_ids, Tensor! expert_offsets, "
" Tensor! problem_sizes1, Tensor! problem_sizes2, "
" Tensor! input_permutation, "
" Tensor! output_permutation, int num_experts, "
" int n, int k, Tensor? blockscale_offsets, "
" bool is_gated) -> ()");
ops.impl("get_cutlass_moe_mm_data", torch::kCUDA, &get_cutlass_moe_mm_data);
// compute per-expert problem sizes from expert_first_token_offset
// produced by vLLM's moe_permute kernel
ops.def(
"get_cutlass_moe_mm_problem_sizes_from_expert_offsets("
" Tensor expert_first_token_offset, "
" Tensor! problem_sizes1, "
" Tensor! problem_sizes2, "
" int n, int k, bool swap_ab) -> ()");
ops.impl("get_cutlass_moe_mm_problem_sizes_from_expert_offsets", torch::kCUDA,
&get_cutlass_moe_mm_problem_sizes_from_expert_offsets);
// A function that computes data required to run fused MoE with w8a8 grouped
// GEMM in batched expert format. It takes expert_num_tokens
// 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_batched_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) -> ()");
ops.impl("get_cutlass_batched_moe_mm_data", torch::kCUDA,
&get_cutlass_batched_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) -> "
"bool");
ops.impl("cutlass_scaled_mm_supports_block_fp8",
&cutlass_scaled_mm_supports_block_fp8);
// SM100 CUTLASS MLA decode
ops.def(
"sm100_cutlass_mla_decode(Tensor! out, Tensor! lse, Tensor q_nope,"
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment