mirror of
https://github.com/NVIDIA/TensorRT-LLM.git
synced 2026-01-14 06:27:45 +08:00
b3e6723dbc
* Adding FP8 BMM from Codegen Signed-off-by: Olya Kozlova <okozlova@s4124-0110.nvidia.com> * Fixed licenses Signed-off-by: Olya Kozlova <okozlova@s4124-0062.nvidia.com> --------- Signed-off-by: Olya Kozlova <okozlova@s4124-0110.nvidia.com> Signed-off-by: Olya Kozlova <okozlova@s4124-0062.nvidia.com> Co-authored-by: Olya Kozlova <okozlova@6u1g-0018.nvidia.com> Co-authored-by: Olya Kozlova <okozlova@s4124-0062.nvidia.com>
112 lines
3.6 KiB
C++
112 lines
3.6 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.
|
|
*/
|
|
|
|
#pragma once
|
|
|
|
#include "tensorrt_llm/common/cudaDriverWrapper.h"
|
|
#include "tensorrt_llm/kernels/multiHeadAttentionCommon.h"
|
|
#include "tensorrt_llm/kernels/trtllmGenKernels/batchedGemm/kernelList.h"
|
|
|
|
namespace tensorrt_llm::kernels
|
|
{
|
|
class TrtllmGenBatchedGemmRunner
|
|
{
|
|
public:
|
|
explicit TrtllmGenBatchedGemmRunner(
|
|
Data_type outputType, int64_t gemmBatchSize, int64_t tileSize, bool useDeepSeekFp8, bool batchModeM);
|
|
|
|
void run(int32_t m, int32_t n, int32_t k, void* a, void* b, void* c, float const* cScale, float* dDqSfsA,
|
|
float* dDqSfsB, float* dDqSfsC, std::vector<int32_t> const& batchedM, std::vector<int32_t> const& batchedN,
|
|
CUstream stream);
|
|
|
|
private:
|
|
Data_type mOutputType;
|
|
int32_t mGemmBatchSize;
|
|
bool mUseDeepSeekFp8;
|
|
BatchMode mBatchMode;
|
|
bool mBatchModeM;
|
|
int32_t mTileSize;
|
|
TrtllmGenBatchedStridedGemmInfo const* mKernelInfo;
|
|
std::shared_ptr<tensorrt_llm::common::CUDADriverWrapper> mDriver;
|
|
CUmodule mModule;
|
|
CUfunction mFunction;
|
|
|
|
static std::array<int, 16> const srcToDstBlk16RowMap;
|
|
static std::array<int, 32> const srcToDstBlk32RowMap;
|
|
|
|
template <Data_type T>
|
|
void shuffleMatrixA(void const* input, void* output, int B, int M, int K, int epilogueTileM);
|
|
};
|
|
|
|
// clang-format off
|
|
inline const std::array<int, 16> TrtllmGenBatchedGemmRunner::srcToDstBlk16RowMap =
|
|
{
|
|
0, 8,
|
|
1, 9,
|
|
2, 10,
|
|
3, 11,
|
|
4, 12,
|
|
5, 13,
|
|
6, 14,
|
|
7, 15
|
|
};
|
|
|
|
inline const std::array<int, 32> TrtllmGenBatchedGemmRunner::srcToDstBlk32RowMap =
|
|
{
|
|
0, 8, 16, 24,
|
|
1, 9, 17, 25,
|
|
2, 10, 18, 26,
|
|
3, 11, 19, 27,
|
|
4, 12, 20, 28,
|
|
5, 13, 21, 29,
|
|
6, 14, 22, 30,
|
|
7, 15, 23, 31
|
|
};
|
|
|
|
// clang-format on
|
|
|
|
template <Data_type T>
|
|
void TrtllmGenBatchedGemmRunner::shuffleMatrixA(void const* input, void* output, int B, int M, int K, int epilogueTileM)
|
|
{
|
|
int shuffleBlockSize = 16;
|
|
if (epilogueTileM % 128 == 0)
|
|
{
|
|
shuffleBlockSize = 32;
|
|
}
|
|
int numBytesPerRow = K * get_size_in_bytes(T);
|
|
|
|
std::vector<uint8_t> tmp(M * numBytesPerRow);
|
|
for (int batchIndex = 0; batchIndex < B; ++batchIndex)
|
|
{
|
|
int const batchRowStride = batchIndex * M;
|
|
for (int mi = 0; mi < M; ++mi)
|
|
{
|
|
int const dstRowBlockIdx = mi / shuffleBlockSize;
|
|
int const srcRowInBlockIdx = mi % shuffleBlockSize;
|
|
int const dstRowInBlockIdx = shuffleBlockSize == 16 ? srcToDstBlk16RowMap[srcRowInBlockIdx]
|
|
: srcToDstBlk32RowMap[srcRowInBlockIdx];
|
|
int const dstRowIdx = dstRowBlockIdx * shuffleBlockSize + dstRowInBlockIdx;
|
|
std::memcpy(&tmp[dstRowIdx * numBytesPerRow],
|
|
&reinterpret_cast<uint8_t const*>(input)[(batchRowStride + mi) * numBytesPerRow], numBytesPerRow);
|
|
}
|
|
|
|
// Copy tmp data to the output pointer.
|
|
std::memcpy(
|
|
&reinterpret_cast<uint8_t*>(output)[batchRowStride * numBytesPerRow], tmp.data(), M * numBytesPerRow);
|
|
}
|
|
}
|
|
} // namespace tensorrt_llm::kernels
|