TensorRT-LLMs/cpp/tensorrt_llm/thop/fp4Gemm.cpp
Yihan Wang 9df4dad3b6
[None][fix] Introduce inline namespace to avoid symbol collision (#9541)
Signed-off-by: Yihan Wang <yihwang@nvidia.com>
2025-12-12 23:32:15 +08:00

337 lines
13 KiB
C++

/*
* Copyright (c) 2020-2023, NVIDIA CORPORATION. All rights reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
#include "cutlass_extensions/gemm_configs.h"
#include "tensorrt_llm/common/cudaUtils.h"
#include "tensorrt_llm/kernels/quantization.h"
#include "tensorrt_llm/thop/thUtils.h"
#include "tensorrt_llm/thop/userbuffersTensor.h"
#if defined(USING_OSS_CUTLASS_FP4_GEMM)
#include "tensorrt_llm/kernels/cutlass_kernels/include/fp4_gemm.h"
#else
#include "fp4_gemm.h"
#endif
#include <ATen/cuda/EmptyTensor.h>
#include <ATen/native/cuda/Resize.h>
#include <cstddef>
#include <cuda_fp16.h>
#include <cstdint>
#include <functional>
#include <type_traits>
#include <vector>
namespace tkc = tensorrt_llm::cutlass_extensions;
#if defined(USING_OSS_CUTLASS_FP4_GEMM)
using tensorrt_llm::kernels::cutlass_kernels::FP4GemmType;
using tensorrt_llm::kernels::cutlass_kernels::CutlassFp4GemmRunner;
using tensorrt_llm::kernels::cutlass_kernels::CutlassFp4GemmRunnerInterface;
#else
using tensorrt_llm::kernels::internal_cutlass_kernels::FP4GemmType;
using tensorrt_llm::kernels::internal_cutlass_kernels::CutlassFp4GemmRunner;
using tensorrt_llm::kernels::internal_cutlass_kernels::CutlassFp4GemmRunnerInterface;
#endif
TRTLLM_NAMESPACE_BEGIN
namespace torch_ext
{
namespace
{
tkc::CutlassGemmConfig getDefaultGemmConfig(int64_t m, int64_t n, int64_t k, FP4GemmType fp4GemmType)
{
auto const sm = tensorrt_llm::common::getSMVersion();
if (fp4GemmType == FP4GemmType::W4A4_NVFP4_NVFP4)
{
if (sm >= 120)
{
return tkc::CutlassGemmConfig(tkc::CutlassTileConfigSM120::CtaShape128x128x256B,
tkc::MainloopScheduleType::AUTO, tkc::EpilogueScheduleType::AUTO,
tkc::ClusterShape::ClusterShape_1x1x1);
}
else if (sm == 103)
{
return tkc::CutlassGemmConfig(tkc::CutlassTileConfigSM100::CtaShape128x256x256B,
tkc::MainloopScheduleType::AUTO, tkc::EpilogueScheduleType::AUTO,
tkc::ClusterShape::ClusterShape_1x1x1);
}
else
{
return tkc::CutlassGemmConfig(tkc::CutlassTileConfigSM100::CtaShape128x256x128B,
tkc::MainloopScheduleType::AUTO, tkc::EpilogueScheduleType::AUTO,
tkc::ClusterShape::ClusterShape_1x1x1);
}
}
else if (fp4GemmType == FP4GemmType::W4A8_MXFP4_MXFP8)
{
return tkc::CutlassGemmConfig(tkc::CutlassTileConfigSM100::CtaShape128x256x256B,
tkc::MainloopScheduleType::AUTO, tkc::EpilogueScheduleType::AUTO, tkc::ClusterShape::ClusterShape_4x4x1);
}
else
{
return tkc::CutlassGemmConfig(tkc::CutlassTileConfigSM100::CtaShape128x256x128B,
tkc::MainloopScheduleType::AUTO, tkc::EpilogueScheduleType::AUTO, tkc::ClusterShape::ClusterShape_1x1x1);
}
}
template <typename T>
void runGemm(at::Tensor& out, at::Tensor const& mat1, at::Tensor const& mat2, at::Tensor const& mat1Scale,
at::Tensor const& mat2Scale, at::Tensor const& globalScale, int64_t m, int64_t n, int64_t k, int64_t batch_count,
tkc::CutlassGemmConfig const& gemmConfig, FP4GemmType fp4GemmType)
{
if (fp4GemmType == FP4GemmType::W4A8_MXFP4_MXFP8)
{
CutlassFp4GemmRunner<T, FP4GemmType::W4A8_MXFP4_MXFP8> gemmRunner;
int64_t const wsBytes = gemmRunner.getWorkspaceSize(m, n, k, batch_count);
at::Tensor workspace = at::detail::empty_cuda({wsBytes}, at::ScalarType::Char, mat1.device(), std::nullopt);
gemmRunner.gemm(out.data_ptr(), mat1.const_data_ptr(), mat2.const_data_ptr(), mat1Scale.const_data_ptr(),
mat2Scale.const_data_ptr(), globalScale.data_ptr<float>(), m, n, k, batch_count, gemmConfig,
reinterpret_cast<char*>(workspace.data_ptr()), wsBytes, at::cuda::getCurrentCUDAStream(mat1.get_device()));
}
else if (fp4GemmType == FP4GemmType::W4A4_NVFP4_NVFP4)
{
CutlassFp4GemmRunner<T, FP4GemmType::W4A4_NVFP4_NVFP4> gemmRunner;
int64_t const wsBytes = gemmRunner.getWorkspaceSize(m, n, k, batch_count);
at::Tensor workspace = at::detail::empty_cuda({wsBytes}, at::ScalarType::Char, mat1.device(), std::nullopt);
gemmRunner.gemm(out.data_ptr(), mat1.const_data_ptr(), mat2.const_data_ptr(), mat1Scale.const_data_ptr(),
mat2Scale.const_data_ptr(), globalScale.data_ptr<float>(), m, n, k, batch_count, gemmConfig,
reinterpret_cast<char*>(workspace.data_ptr()), wsBytes, at::cuda::getCurrentCUDAStream(mat1.get_device()));
}
}
// mat1: [B, M, K / 2], FLOAT4_E2M1X2 or [B, M, K], FLOAT8_E4M3FN
// mat2: [B, N, K / 2], FLOAT4_E2M1X2
// out: [B, M, N], fp16/bf16/fp32
// mat1Scale: ceil(M / 128) * 128 * ceil(K / sfVecSize / 4) * 4, SF_DTYPE (UE4M3 or UE8M0)
// mat2Scale: ceil(N / 128) * 128 * ceil(K / sfVecSize / 4) * 4, SF_DTYPE (UE4M3 or UE8M0)
// globalScale: [1], 1 / (((448 * 6) / mat1.abs().max()) * ((448 * 6) / mat2.abs().max()))
// B = 1 for GEMM op as a special case
// Only W4A4_NVFP4 and W4A8_MXFP4_FP8 are currently supported
at::Tensor fp4_bmm_impl(at::Tensor const& mat1, at::Tensor const& mat2, at::Tensor const& mat1Scale,
at::Tensor const& mat2Scale, at::Tensor const& globalScale, FP4GemmType fp4GemmType,
std::optional<c10::ScalarType> out_dtype, bool to_userbuffers = false,
tkc::CutlassGemmConfig const* maybe_config = nullptr)
{
if (fp4GemmType == FP4GemmType::W4A8_MXFP4_MXFP8)
{
CHECK_INPUT(mat1, torch::kFloat8_e4m3fn);
CHECK_INPUT(mat2, FLOAT4_E2M1X2);
}
else if (fp4GemmType == FP4GemmType::W4A4_NVFP4_NVFP4)
{
CHECK_INPUT(mat1, FLOAT4_E2M1X2);
CHECK_INPUT(mat2, FLOAT4_E2M1X2);
}
int mat2_k_scale = fp4GemmType == FP4GemmType::W4A8_MXFP4_MXFP8 ? 2 : 1;
CHECK_INPUT(mat1Scale, SF_DTYPE);
CHECK_INPUT(mat2Scale, SF_DTYPE);
CHECK_INPUT(globalScale, at::ScalarType::Float);
int64_t m, n, k, b;
if (mat1.dim() == 2)
{
TORCH_CHECK(mat2.dim() == 2, "mat2 must be a matrix");
TORCH_CHECK(mat1.sizes()[1] == mat2.sizes()[1] * mat2_k_scale, "mat1 and mat2 shapes cannot be multiplied (",
mat1.sizes()[0], "x", mat1.sizes()[1], " and ", mat2.sizes()[0], "x", mat2.sizes()[1], ")");
m = mat1.sizes()[0];
n = mat2.sizes()[0];
k = mat2.sizes()[1] * 2;
b = 1;
}
else if (mat1.dim() == 3)
{
TORCH_CHECK(mat2.dim() == 3, "mat2 must be a batch of matrices");
TORCH_CHECK(mat1.sizes()[0] == mat2.sizes()[0], "mat1 and mat2 must have the same batch size (",
mat1.sizes()[0], " and ", mat2.sizes()[0], ")");
TORCH_CHECK(mat1.sizes()[2] == mat2.sizes()[2] * mat2_k_scale, "mat1 and mat2 shapes cannot be multiplied (",
mat1.sizes()[1], "x", mat1.sizes()[2], " and ", mat2.sizes()[1], "x", mat2.sizes()[2], ")");
m = mat1.sizes()[1];
n = mat2.sizes()[1];
k = mat2.sizes()[2] * 2;
b = mat1.sizes()[0];
}
else
{
C10_THROW_ERROR(NotImplementedError, "mat1 must be a matrix or a batch of matrices");
}
auto config = maybe_config ? *maybe_config : getDefaultGemmConfig(m, n, k, fp4GemmType);
constexpr int alignment = 32;
TORCH_CHECK(k % alignment == 0, "Expected k to be divisible by ", alignment, ", but got mat1 shape: (",
mat1.sizes()[0], "x", mat1.sizes()[1], "), k: ", k, ".");
TORCH_CHECK(n % alignment == 0, "Expected n to be divisible by ", alignment, ", but got mat2 shape: (",
mat2.sizes()[0], "x", mat2.sizes()[1], ").");
if (!out_dtype)
{
out_dtype = torch::kHalf;
}
TORCH_CHECK(out_dtype == torch::kFloat || out_dtype == torch::kHalf || out_dtype == torch::kBFloat16,
"out_dtype must be one of fp16/bf16/fp32. It defaults to fp16.");
std::vector<int64_t> out_shape = mat1.dim() == 2 ? std::vector<int64_t>{m, n} : std::vector<int64_t>{b, m, n};
at::Tensor out;
if (to_userbuffers)
{
out = torch_ext::create_userbuffers_tensor(out_shape, out_dtype.value()).first;
}
else
{
out = at::detail::empty_cuda(out_shape, out_dtype.value(), mat1.device(), std::nullopt);
}
switch (out_dtype.value())
{
case at::ScalarType::Half:
runGemm<half>(out, mat1, mat2, mat1Scale, mat2Scale, globalScale, m, n, k, b, config, fp4GemmType);
break;
case at::ScalarType::BFloat16:
runGemm<__nv_bfloat16>(out, mat1, mat2, mat1Scale, mat2Scale, globalScale, m, n, k, b, config, fp4GemmType);
break;
case at::ScalarType::Float:
runGemm<float>(out, mat1, mat2, mat1Scale, mat2Scale, globalScale, m, n, k, b, config, fp4GemmType);
break;
default: C10_THROW_ERROR(NotImplementedError, "out_dtype must be one of fp16/bf16/fp32.");
}
return out;
}
} // namespace
at::Tensor fp4_bmm(at::Tensor const& mat1, at::Tensor const& mat2, at::Tensor const& mat1Scale,
at::Tensor const& mat2Scale, at::Tensor const& globalScale, int64_t fp4GemmType,
std::optional<c10::ScalarType> out_dtype)
{
// The functional version of this op does not do any profiling; use the profiler class below instead for
// better performance.
// Note that we can still add a heuristic here.
return fp4_bmm_impl(
mat1, mat2, mat1Scale, mat2Scale, globalScale, static_cast<FP4GemmType>(fp4GemmType), out_dtype);
}
class FP4GemmRunner : public torch::CustomClassHolder
{
public:
explicit FP4GemmRunner(at::ScalarType outputDtype, int64_t fp4GemmType)
: mOutputDtype(outputDtype)
, mfp4GemmType(static_cast<FP4GemmType>(fp4GemmType))
{
if (outputDtype == at::ScalarType::Half)
{
if (mfp4GemmType == FP4GemmType::W4A8_MXFP4_MXFP8)
{
mGemmRunner = std::make_unique<CutlassFp4GemmRunner<half, FP4GemmType::W4A8_MXFP4_MXFP8>>();
}
else if (mfp4GemmType == FP4GemmType::W4A4_NVFP4_NVFP4)
{
mGemmRunner = std::make_unique<CutlassFp4GemmRunner<half, FP4GemmType::W4A4_NVFP4_NVFP4>>();
}
}
else if (outputDtype == at::ScalarType::Float)
{
if (mfp4GemmType == FP4GemmType::W4A8_MXFP4_MXFP8)
{
mGemmRunner = std::make_unique<CutlassFp4GemmRunner<float, FP4GemmType::W4A8_MXFP4_MXFP8>>();
}
else if (mfp4GemmType == FP4GemmType::W4A4_NVFP4_NVFP4)
{
mGemmRunner = std::make_unique<CutlassFp4GemmRunner<float, FP4GemmType::W4A4_NVFP4_NVFP4>>();
}
}
#ifdef ENABLE_BF16
else if (outputDtype == at::ScalarType::BFloat16)
{
if (mfp4GemmType == FP4GemmType::W4A8_MXFP4_MXFP8)
{
mGemmRunner = std::make_unique<CutlassFp4GemmRunner<__nv_bfloat16, FP4GemmType::W4A8_MXFP4_MXFP8>>();
}
else if (mfp4GemmType == FP4GemmType::W4A4_NVFP4_NVFP4)
{
mGemmRunner = std::make_unique<CutlassFp4GemmRunner<__nv_bfloat16, FP4GemmType::W4A4_NVFP4_NVFP4>>();
}
}
#endif
else
{
C10_THROW_ERROR(NotImplementedError, "out_dtype must be one of fp16/bf16/fp32.");
}
mConfigs = mGemmRunner->getConfigs();
}
at::Tensor runGemm(at::Tensor const& mat1, at::Tensor const& mat2, at::Tensor const& mat1Scale,
at::Tensor const& mat2Scale, at::Tensor const& globalScale, bool to_userbuffers, int64_t configIdx) const
{
tkc::CutlassGemmConfig const* config = nullptr;
if (configIdx != -1)
{
TORCH_CHECK(configIdx >= 0 && configIdx < getNumConfigs());
config = &mConfigs.at(configIdx);
}
return fp4_bmm_impl(
mat1, mat2, mat1Scale, mat2Scale, globalScale, mfp4GemmType, mOutputDtype, to_userbuffers, config);
}
at::ScalarType getOutputDtype() const
{
return mOutputDtype;
}
int64_t getNumConfigs() const
{
return static_cast<int64_t>(mConfigs.size());
}
private:
std::shared_ptr<CutlassFp4GemmRunnerInterface> mGemmRunner{nullptr};
std::vector<tkc::CutlassGemmConfig> mConfigs;
at::ScalarType mOutputDtype;
FP4GemmType mfp4GemmType;
};
} // namespace torch_ext
TRTLLM_NAMESPACE_END
TORCH_LIBRARY_FRAGMENT(trtllm, m)
{
m.class_<tensorrt_llm::torch_ext::FP4GemmRunner>("FP4GemmRunner")
.def(torch::init<at::ScalarType, int64_t>())
.def("run_gemm", &tensorrt_llm::torch_ext::FP4GemmRunner::runGemm)
.def("get_num_configs", &tensorrt_llm::torch_ext::FP4GemmRunner::getNumConfigs);
m.def(
"fp4_bmm(Tensor mat1, Tensor mat2, Tensor mat1Scale, Tensor mat2Scale, Tensor globalScale, int fp4GemmType, "
"ScalarType? out_dtype=None) -> Tensor");
m.def(
"fp4_gemm(Tensor mat1, Tensor mat2, Tensor mat1Scale, Tensor mat2Scale, Tensor globalScale, int fp4GemmType, "
"ScalarType? out_dtype=None) -> Tensor");
}
TORCH_LIBRARY_IMPL(trtllm, CUDA, m)
{
m.impl("fp4_bmm", &tensorrt_llm::torch_ext::fp4_bmm);
m.impl("fp4_gemm", &tensorrt_llm::torch_ext::fp4_bmm);
}