New stable abi cleanup (#46656)

Signed-off-by: Chris Leonard <chleonar@redhat.com>
This commit is contained in:
Chris Leonard
2026-07-03 14:02:26 +08:00
committed by GitHub
parent 2dfaae752b
commit fbc9ba6d30
50 changed files with 96 additions and 44 deletions
@@ -37,6 +37,7 @@ void cutlass_gemm_caller(
typename GemmKernel::MainloopArguments mainloop_args,
typename GemmKernel::EpilogueArguments epilogue_args,
typename GemmKernel::TileSchedulerArguments scheduler = {}) {
const torch::stable::accelerator::DeviceGuard device_guard(device.index());
cutlass::KernelHardwareInfo hw_info;
typename GemmKernel::Arguments args{cutlass::gemm::GemmUniversalMode::kGemm,
prob_shape,
@@ -1,6 +1,6 @@
#include "scaled_mm_kernels.hpp"
#include "scaled_mm_sm90_int8_dispatch.cuh"
#include "cutlass_extensions/epilogue/scaled_mm_epilogues_c3x.hpp"
#include "libtorch_stable/cutlass_extensions/epilogue/scaled_mm_epilogues_c3x.hpp"
namespace vllm {
@@ -1,6 +1,6 @@
#include "scaled_mm_kernels.hpp"
#include "scaled_mm_blockwise_sm100_fp8_dispatch.cuh"
#include "cutlass_extensions/epilogue/scaled_mm_epilogues_c3x.hpp"
#include "libtorch_stable/cutlass_extensions/epilogue/scaled_mm_epilogues_c3x.hpp"
namespace vllm {
@@ -1,6 +1,6 @@
#include "scaled_mm_kernels.hpp"
#include "scaled_mm_blockwise_sm120_fp8_dispatch.cuh"
#include "cutlass_extensions/epilogue/scaled_mm_epilogues_c3x.hpp"
#include "libtorch_stable/cutlass_extensions/epilogue/scaled_mm_epilogues_c3x.hpp"
namespace vllm {
@@ -1,7 +1,7 @@
#include "scaled_mm_kernels.hpp"
#include "scaled_mm_blockwise_sm90_fp8_dispatch.cuh"
#include "cutlass_extensions/epilogue/scaled_mm_epilogues_c3x.hpp"
#include "libtorch_stable/cutlass_extensions/epilogue/scaled_mm_epilogues_c3x.hpp"
namespace vllm {
@@ -4,7 +4,7 @@
#include "scaled_mm.cuh"
#include "cutlass_gemm_caller.cuh"
#include "cutlass_extensions/epilogue/scaled_mm_epilogues_c3x.hpp"
#include "libtorch_stable/cutlass_extensions/epilogue/scaled_mm_epilogues_c3x.hpp"
/**
* This file defines Gemm kernel configurations for SM100 (fp8) based on the
@@ -1,7 +1,7 @@
#include "scaled_mm_kernels.hpp"
#include "scaled_mm_sm120_fp8_dispatch.cuh"
#include "core/batch_invariant.hpp"
#include "cutlass_extensions/epilogue/scaled_mm_epilogues_c3x.hpp"
#include "libtorch_stable/cutlass_extensions/epilogue/scaled_mm_epilogues_c3x.hpp"
namespace vllm {
@@ -4,7 +4,7 @@
#include "scaled_mm.cuh"
#include "cutlass_gemm_caller.cuh"
#include "cutlass_extensions/epilogue/scaled_mm_epilogues_c3x.hpp"
#include "libtorch_stable/cutlass_extensions/epilogue/scaled_mm_epilogues_c3x.hpp"
/**
* This file defines Gemm kernel configurations for SM90 (fp8) based on the Gemm
@@ -1,6 +1,6 @@
#include "scaled_mm_kernels.hpp"
#include "scaled_mm_sm90_int8_dispatch.cuh"
#include "cutlass_extensions/epilogue/scaled_mm_epilogues_c3x.hpp"
#include "libtorch_stable/cutlass_extensions/epilogue/scaled_mm_epilogues_c3x.hpp"
namespace vllm {
@@ -74,6 +74,8 @@ void run_get_group_gemm_starts(
bool per_act_token = a_scales.numel() != 1;
bool per_out_ch = b_scales.numel() != num_experts;
const torch::stable::accelerator::DeviceGuard device_guard(
a_tensors.get_device_index());
auto stream = get_current_cuda_stream(a_tensors.get_device_index());
if (false) {
@@ -7,8 +7,9 @@
#include "cutlass/gemm/device/gemm_universal_adapter.h"
#include <torch/csrc/stable/ops.h>
#include "cutlass_extensions/epilogue/scaled_mm_epilogues_c3x.hpp"
#include "libtorch_stable/cutlass_extensions/epilogue/scaled_mm_epilogues_c3x.hpp"
#include "libtorch_stable/cutlass_extensions/common.hpp"
#include "libtorch_stable/torch_utils.h"
#include "get_group_starts.cuh"
using namespace cute;
@@ -103,6 +104,8 @@ void cutlass_group_gemm_caller(torch::stable::Tensor& out_tensors,
int num_experts = static_cast<int>(expert_offsets.size(0));
const torch::stable::accelerator::DeviceGuard device_guard(
a_tensors.get_device_index());
auto stream = get_current_cuda_stream(a_tensors.get_device_index());
auto device = a_tensors.device();
@@ -212,6 +212,8 @@ void get_cutlass_moe_mm_problem_sizes_from_expert_offsets_caller(
"n and k must fit in int32");
int const num_experts = static_cast<int>(num_experts64);
const torch::stable::accelerator::DeviceGuard device_guard(
expert_first_token_offset.get_device_index());
auto stream =
get_current_cuda_stream(expert_first_token_offset.get_device_index());
@@ -241,6 +243,7 @@ void get_cutlass_moe_mm_data_caller(
const std::optional<torch::stable::Tensor>& blockscale_offsets,
const bool is_gated) {
auto device = topk_ids.device();
const torch::stable::accelerator::DeviceGuard device_guard(device.index());
auto stream = get_current_cuda_stream(device.index());
torch::stable::Tensor atomic_buffer = torch::stable::new_zeros(
topk_ids, {num_experts}, torch::headeronly::ScalarType::Int);
@@ -311,6 +314,8 @@ void get_cutlass_batched_moe_mm_data_caller(
const torch::stable::Tensor& expert_num_tokens,
const int64_t num_local_experts, const int64_t padded_m, const int64_t n,
const int64_t k) {
const torch::stable::accelerator::DeviceGuard device_guard(
expert_offsets.get_device_index());
auto stream = get_current_cuda_stream(expert_offsets.get_device_index());
if (num_local_experts * padded_m > SWAP_AB_THRESHOLD) {
@@ -156,6 +156,7 @@ inline void cutlass_gemm_caller(torch::stable::Tensor& out,
torch::stable::empty(workspace_size, torch::headeronly::ScalarType::Byte,
std::nullopt, device);
const torch::stable::accelerator::DeviceGuard device_guard(device.index());
auto stream = get_current_cuda_stream(device.index());
CUTLASS_CHECK(gemm_op.can_implement(args));
@@ -1,6 +1,6 @@
#include "../../../../quantization/w8a8/fp8/common.cuh"
#include "../../../dispatch_utils.h"
#include "../../../../cub_helpers.h"
#include "../../../cub_helpers.h"
#include "../../vectorization_utils.cuh"
#include "../../../torch_utils.h"
#include <torch/csrc/stable/macros.h>
@@ -203,6 +203,8 @@ void per_token_group_quant_8bit(const torch::stable::Tensor& input,
STD_TORCH_CHECK(input.numel() % group_size == 0);
STD_TORCH_CHECK(output_s.dim() == 2);
const torch::stable::accelerator::DeviceGuard device_guard(
input.get_device_index());
cudaStream_t stream = get_current_cuda_stream();
constexpr int THREADS_PER_GROUP = 16;
@@ -506,6 +508,8 @@ void per_token_group_quant_8bit_packed(const torch::stable::Tensor& input,
"]; got [", output_s_packed.stride(0), ", ",
output_s_packed.stride(1), "].");
const torch::stable::accelerator::DeviceGuard device_guard(
input.get_device_index());
cudaStream_t stream = get_current_cuda_stream();
constexpr int THREADS_PER_GROUP = 8;
@@ -5,7 +5,7 @@
#include "../../../dispatch_utils.h"
#include "../../../torch_utils.h"
#include "../../vectorization_utils.cuh"
#include "../../../../cub_helpers.h"
#include "../../../cub_helpers.h"
static inline __device__ int8_t float_to_int8_rn(float x) {
#ifdef USE_ROCM