mirror of
https://github.com/NVIDIA/TensorRT-LLM.git
synced 2026-01-14 06:27:45 +08:00
c89653021e
* Update TensorRT-LLM --------- Co-authored-by: Eddie-Wang1120 <81598289+Eddie-Wang1120@users.noreply.github.com> Co-authored-by: Shixiaowei02 <39303645+Shixiaowei02@users.noreply.github.com>
252 lines
12 KiB
Plaintext
252 lines
12 KiB
Plaintext
/*
|
|
* SPDX-FileCopyrightText: Copyright (c) 1993-2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
|
* SPDX-License-Identifier: Apache-2.0
|
|
*
|
|
* 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 "kvCacheUpdateKernels.h"
|
|
|
|
#include <array>
|
|
|
|
#include "tensorrt_llm/common/cudaUtils.h"
|
|
#include "tensorrt_llm/kernels/kvCacheUtils.h"
|
|
|
|
namespace tensorrt_llm::kernels::parallel_decoding
|
|
{
|
|
|
|
static constexpr int kUpdateKVCacheKernelShmSize = 16384;
|
|
|
|
template <typename KVCacheBuffer, int MaxLayerCount, typename MoveEltType>
|
|
__global__ void updateKVCacheDraftTokenLocationBatchedKernel(std::array<KVCacheBuffer, MaxLayerCount> kvCacheBuffers,
|
|
const int* seqAcceptedDraftTokenOffsets, const IndexType* packedAcceptedDraftTokensIndices,
|
|
const int32_t* pastKeyValueLengths, int rewindDraftTokenCount, int eltCountPerHead)
|
|
{
|
|
int seqIdx = blockIdx.x;
|
|
int headIdx = blockIdx.y;
|
|
int layerIdx = blockIdx.z;
|
|
int warpIdx = threadIdx.x / 32;
|
|
int warpCount = blockDim.x / 32;
|
|
int laneIdx = threadIdx.x & 0x1f;
|
|
int seqDraftTokenStart = seqAcceptedDraftTokenOffsets[seqIdx];
|
|
int seqDraftTokenEnd = seqAcceptedDraftTokenOffsets[seqIdx + 1];
|
|
int seqDraftCount = seqDraftTokenEnd - seqDraftTokenStart;
|
|
if (seqDraftCount == 0)
|
|
{
|
|
return;
|
|
}
|
|
KVCacheBuffer& kvCacheBuffer = kvCacheBuffers[layerIdx];
|
|
int tokenStartIdx = pastKeyValueLengths[seqIdx] - rewindDraftTokenCount;
|
|
int maxEltCountPerMove = kUpdateKVCacheKernelShmSize / sizeof(MoveEltType) / seqDraftCount;
|
|
int eltCountPerMove = min(maxEltCountPerMove, eltCountPerHead);
|
|
__shared__ char loadSmemBuffer[kUpdateKVCacheKernelShmSize];
|
|
auto* eltLoadSmemBuffer = reinterpret_cast<MoveEltType*>(&loadSmemBuffer[0]);
|
|
for (int startChannelOffset = 0; startChannelOffset < eltCountPerHead; startChannelOffset += eltCountPerMove)
|
|
{
|
|
int eltCountCurrentMove = min(eltCountPerMove, eltCountPerHead - startChannelOffset);
|
|
// load K
|
|
for (int tokenIdx = warpIdx; tokenIdx < seqDraftCount; tokenIdx += warpCount)
|
|
{
|
|
int tokenPos = packedAcceptedDraftTokensIndices[seqDraftTokenStart + tokenIdx];
|
|
auto* tokenSmemBuffer = eltLoadSmemBuffer + tokenIdx * eltCountCurrentMove;
|
|
int tokenKVPosition = tokenStartIdx + tokenPos;
|
|
auto* kPtr = reinterpret_cast<MoveEltType*>(kvCacheBuffer.getKBlockPtr(seqIdx, tokenKVPosition));
|
|
for (int loadChannelIdx = laneIdx; loadChannelIdx < eltCountCurrentMove; loadChannelIdx += 32)
|
|
{
|
|
int channelIdx = loadChannelIdx + startChannelOffset;
|
|
int kvLocationIdx = kvCacheBuffer.getKVLocalIdx(tokenKVPosition, headIdx, eltCountPerHead, channelIdx);
|
|
tokenSmemBuffer[loadChannelIdx] = kPtr[kvLocationIdx];
|
|
}
|
|
}
|
|
__syncthreads();
|
|
// store K
|
|
for (int tokenIdx = warpIdx; tokenIdx < seqDraftCount; tokenIdx += warpCount)
|
|
{
|
|
int tokenPos = tokenIdx;
|
|
auto* tokenSmemBuffer = eltLoadSmemBuffer + tokenIdx * eltCountCurrentMove;
|
|
int tokenKVPosition = tokenStartIdx + tokenPos;
|
|
auto* kPtr = reinterpret_cast<MoveEltType*>(kvCacheBuffer.getKBlockPtr(seqIdx, tokenKVPosition));
|
|
for (int loadChannelIdx = laneIdx; loadChannelIdx < eltCountCurrentMove; loadChannelIdx += 32)
|
|
{
|
|
int channelIdx = loadChannelIdx + startChannelOffset;
|
|
int kvLocationIdx = kvCacheBuffer.getKVLocalIdx(tokenKVPosition, headIdx, eltCountPerHead, channelIdx);
|
|
kPtr[kvLocationIdx] = tokenSmemBuffer[loadChannelIdx];
|
|
}
|
|
}
|
|
__syncthreads();
|
|
// load V
|
|
for (int tokenIdx = warpIdx; tokenIdx < seqDraftCount; tokenIdx += warpCount)
|
|
{
|
|
int tokenPos = packedAcceptedDraftTokensIndices[seqDraftTokenStart + tokenIdx];
|
|
auto* tokenSmemBuffer = eltLoadSmemBuffer + tokenIdx * eltCountCurrentMove;
|
|
int tokenKVPosition = tokenStartIdx + tokenPos;
|
|
auto* vPtr = reinterpret_cast<MoveEltType*>(kvCacheBuffer.getVBlockPtr(seqIdx, tokenKVPosition));
|
|
for (int loadChannelIdx = laneIdx; loadChannelIdx < eltCountCurrentMove; loadChannelIdx += 32)
|
|
{
|
|
int channelIdx = loadChannelIdx + startChannelOffset;
|
|
int kvLocationIdx = kvCacheBuffer.getKVLocalIdx(tokenKVPosition, headIdx, eltCountPerHead, channelIdx);
|
|
tokenSmemBuffer[loadChannelIdx] = vPtr[kvLocationIdx];
|
|
}
|
|
}
|
|
__syncthreads();
|
|
// store V
|
|
for (int tokenIdx = warpIdx; tokenIdx < seqDraftCount; tokenIdx += warpCount)
|
|
{
|
|
int tokenPos = tokenIdx;
|
|
auto* tokenSmemBuffer = eltLoadSmemBuffer + tokenPos * eltCountCurrentMove;
|
|
int tokenKVPosition = tokenStartIdx + tokenPos;
|
|
auto* vPtr = reinterpret_cast<MoveEltType*>(kvCacheBuffer.getVBlockPtr(seqIdx, tokenKVPosition));
|
|
for (int loadChannelIdx = laneIdx; loadChannelIdx < eltCountCurrentMove; loadChannelIdx += 32)
|
|
{
|
|
int channelIdx = loadChannelIdx + startChannelOffset;
|
|
int kvLocationIdx = kvCacheBuffer.getKVLocalIdx(tokenKVPosition, headIdx, eltCountPerHead, channelIdx);
|
|
vPtr[kvLocationIdx] = tokenSmemBuffer[loadChannelIdx];
|
|
}
|
|
}
|
|
__syncthreads();
|
|
}
|
|
}
|
|
|
|
template <typename KVCacheBuffer, int MaxLayerCount>
|
|
void updateKVCacheDraftTokenLocationBatched(const KVCacheBuffer* kvCacheBuffers,
|
|
const int* seqAcceptedDraftTokenOffsets, const IndexType* packedAcceptedDraftTokensIndices,
|
|
const int32_t* pastKeyValueLengths, int layerCount, int seqCount, int numKVHeads, int sizeInBytesPerKVHead,
|
|
int rewindDraftTokenCount, cudaStream_t stream)
|
|
{
|
|
// make sure launch buffer is enough
|
|
static_assert(MaxLayerCount * sizeof(KVCacheBuffer) <= 3072);
|
|
if (seqCount == 0 || layerCount == 0)
|
|
{
|
|
return;
|
|
}
|
|
int alignedBytes = 16;
|
|
while (alignedBytes > 0 && (sizeInBytesPerKVHead % alignedBytes != 0))
|
|
{
|
|
alignedBytes >>= 1;
|
|
}
|
|
TLLM_CHECK_WITH_INFO(alignedBytes > 0, "alignedByte should be positive");
|
|
int eltCountPerHead = sizeInBytesPerKVHead / alignedBytes;
|
|
dim3 grid(seqCount, numKVHeads, layerCount);
|
|
dim3 block(128, 1, 1);
|
|
std::array<KVCacheBuffer, MaxLayerCount> kvCacheBufferArray;
|
|
for (int i = 0; i < layerCount; i++)
|
|
{
|
|
kvCacheBufferArray[i] = kvCacheBuffers[i];
|
|
}
|
|
void (*pKernelFunc)(
|
|
std::array<KVCacheBuffer, MaxLayerCount>, const int*, const IndexType*, const int32_t*, int, int)
|
|
= nullptr;
|
|
switch (alignedBytes)
|
|
{
|
|
case 16:
|
|
{
|
|
pKernelFunc = &updateKVCacheDraftTokenLocationBatchedKernel<KVCacheBuffer, MaxLayerCount, int4>;
|
|
break;
|
|
}
|
|
case 8:
|
|
{
|
|
pKernelFunc = &updateKVCacheDraftTokenLocationBatchedKernel<KVCacheBuffer, MaxLayerCount, int64_t>;
|
|
break;
|
|
}
|
|
case 4:
|
|
{
|
|
pKernelFunc = &updateKVCacheDraftTokenLocationBatchedKernel<KVCacheBuffer, MaxLayerCount, int32_t>;
|
|
break;
|
|
}
|
|
case 2:
|
|
{
|
|
pKernelFunc = &updateKVCacheDraftTokenLocationBatchedKernel<KVCacheBuffer, MaxLayerCount, int16_t>;
|
|
break;
|
|
}
|
|
default:
|
|
{
|
|
TLLM_CHECK_WITH_INFO(alignedBytes == 1, "Strange alignedBytes");
|
|
pKernelFunc = &updateKVCacheDraftTokenLocationBatchedKernel<KVCacheBuffer, MaxLayerCount, int8_t>;
|
|
break;
|
|
}
|
|
}
|
|
pKernelFunc<<<grid, block, 0, stream>>>(kvCacheBufferArray, seqAcceptedDraftTokenOffsets,
|
|
packedAcceptedDraftTokensIndices, pastKeyValueLengths, rewindDraftTokenCount, eltCountPerHead);
|
|
TLLM_CUDA_CHECK(cudaGetLastError());
|
|
}
|
|
|
|
/*!
|
|
* Update KV cache for parallel decoding algorithms.
|
|
* In following examples, we assume we have 2 sequences, accepted count is [3, 2]
|
|
* @tparam KVCacheBuffer : Type of KV cache, should be LinearKVCache or KVBlockArray
|
|
* @param kvCacheBuffers : list of KVCacheBuffer object
|
|
* @param seqAcceptedDraftTokenOffsets : Array of length seqCount + 1, like [0, 3, 5]
|
|
* @param packedAcceptedDraftTokensIndices : Array of length seqAcceptedDraftTokenOffsets[seqCount], each value is in
|
|
* range [0, maxDraftTokenCount - 1]
|
|
* @param pastKeyValueLengths : Array of length seqCount, meaning how many tokens are already in KV cache
|
|
* @param seqCount : Count of sequence
|
|
* @param numKVHeads : Number of KV heads
|
|
* @param sizeInBytesPerKVHead :
|
|
* @param rewindDraftTokenCount
|
|
* @param stream
|
|
*/
|
|
template <typename KVCacheBuffer>
|
|
void updateKVCacheDraftTokenLocation(const KVCacheBuffer* kvCacheBuffers, const int* seqAcceptedDraftTokenOffsets,
|
|
const IndexType* packedAcceptedDraftTokensIndices, const int32_t* pastKeyValueLengths, int layerCount, int seqCount,
|
|
int numKVHeads, int sizeInBytesPerKVHead, int rewindDraftTokenCount, cudaStream_t stream)
|
|
{
|
|
int startLayer = 0;
|
|
static constexpr int kMaxLayersPerIter = 32;
|
|
while (startLayer < layerCount)
|
|
{
|
|
int microBatchLayerCount = std::min(layerCount - startLayer, kMaxLayersPerIter);
|
|
updateKVCacheDraftTokenLocationBatched<KVCacheBuffer, kMaxLayersPerIter>(kvCacheBuffers + startLayer,
|
|
seqAcceptedDraftTokenOffsets, packedAcceptedDraftTokensIndices, pastKeyValueLengths, microBatchLayerCount,
|
|
seqCount, numKVHeads, sizeInBytesPerKVHead, rewindDraftTokenCount, stream);
|
|
startLayer += microBatchLayerCount;
|
|
}
|
|
}
|
|
|
|
void updateLinearKVCacheDraftTokenLocation(const int* seqAcceptedDraftTokenOffsets,
|
|
const IndexType* packedAcceptedDraftTokensIndices, const int32_t* pastKeyValueLengths,
|
|
int8_t* const* pastKeyValueList, int layerCount, int seqCount, int numKVHeads, int sizeInBytesPerKVHead,
|
|
int rewindDraftTokenCount, int maxKVCacheLen, cudaStream_t stream)
|
|
{
|
|
std::vector<KVLinearBuffer> kvLinearBuffers;
|
|
kvLinearBuffers.reserve(layerCount);
|
|
int sizePerToken = numKVHeads * sizeInBytesPerKVHead;
|
|
for (int i = 0; i < layerCount; i++)
|
|
{
|
|
kvLinearBuffers.emplace_back(seqCount, 0, maxKVCacheLen, sizePerToken, maxKVCacheLen, 0, false);
|
|
kvLinearBuffers.back().data = pastKeyValueList[i];
|
|
}
|
|
updateKVCacheDraftTokenLocation(kvLinearBuffers.data(), seqAcceptedDraftTokenOffsets,
|
|
packedAcceptedDraftTokensIndices, pastKeyValueLengths, layerCount, seqCount, numKVHeads, sizeInBytesPerKVHead,
|
|
rewindDraftTokenCount, stream);
|
|
}
|
|
|
|
void updateKVBlockArrayDraftTokenLocation(const int* seqAcceptedDraftTokenOffsets,
|
|
const IndexType* packedAcceptedDraftTokensIndices, const int32_t* pastKeyValueLengths, int64_t* const* pointerArray,
|
|
int layerCount, int seqCount, int numKVHeads, int sizeInBytesPerKVHead, int rewindDraftTokenCount,
|
|
int maxKVCacheLen, int maxBlocksPerSeq, int tokensPerBlock, cudaStream_t stream)
|
|
{
|
|
std::vector<KVBlockArray> kvBlockArrays;
|
|
kvBlockArrays.reserve(layerCount);
|
|
int sizePerToken = numKVHeads * sizeInBytesPerKVHead;
|
|
for (int i = 0; i < layerCount; i++)
|
|
{
|
|
kvBlockArrays.emplace_back(seqCount, maxBlocksPerSeq, tokensPerBlock, sizePerToken, maxKVCacheLen, 0, false);
|
|
kvBlockArrays.back().data = pointerArray[i];
|
|
}
|
|
updateKVCacheDraftTokenLocation(kvBlockArrays.data(), seqAcceptedDraftTokenOffsets,
|
|
packedAcceptedDraftTokensIndices, pastKeyValueLengths, layerCount, seqCount, numKVHeads, sizeInBytesPerKVHead,
|
|
rewindDraftTokenCount, stream);
|
|
}
|
|
|
|
} // namespace tensorrt_llm::kernels::parallel_decoding
|