Skip to content

[None] [feat] Mixed precision context and generation servers #8592

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
wants to merge 7 commits into
base: main
Choose a base branch
from
+560 −47
Draft

[None] [feat] Mixed precision context and generation servers #8592

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
0a98ddf
Add mixed precision context and generation section
timothygao8710 Sep 29, 2025
72c2041
Enhance mixed precision context and generation section
timothygao8710 Sep 30, 2025
e101bfa
Enhance mixed precision section in README
timothygao8710 Oct 6, 2025
1627f0e
Enhance clarity on mixed precision worker characteristics
timothygao8710 Oct 6, 2025
b7d77eb
Change wording
timothygao8710 Oct 7, 2025
ad6c8a0
Merge branch 'main' of https://github.com/NVIDIA/TensorRT-LLM
timothygao8710 Oct 22, 2025
7afe234
Fix commit history
timothygao8710 Oct 22, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
126 changes: 120 additions & 6 deletions cpp/tensorrt_llm/batch_manager/cacheFormatter.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -30,6 +30,7 @@
#include "tensorrt_llm/executor/cache_transmission/agent_utils/connection.h"
#include "tensorrt_llm/executor/cache_transmission/cacheSplitConcat.h"
#include "tensorrt_llm/executor/executor.h"
#include "tensorrt_llm/kernels/unfusedAttentionKernels.h"
#include "tensorrt_llm/runtime/iTensor.h"
#include "tensorrt_llm/runtime/utils/mpiUtils.h"
#include <algorithm>
Expand Down Expand Up @@ -497,6 +498,102 @@ void CacheFormatter::format(tensorrt_llm::batch_manager::TransferSession& sessio
mpi::MpiComm::world().getRank(), "End the sending of KV cache for the request ID:%ld ", llmRequest.mRequestId);
}

namespace {
std::string dataTypeToString(nvinfer1::DataType type)
{
switch (type)
{
case nvinfer1::DataType::kINT64: return "INT64";
case nvinfer1::DataType::kINT32: return "INT32";
case nvinfer1::DataType::kFLOAT: return "FP32";
case nvinfer1::DataType::kBF16: return "BF16";
case nvinfer1::DataType::kHALF: return "FP16";
case nvinfer1::DataType::kINT8: return "INT8";
case nvinfer1::DataType::kUINT8: return "UINT8";
case nvinfer1::DataType::kFP8: return "FP8";
case nvinfer1::DataType::kBOOL: return "BOOL";
default: return "UNKNOWN";
}
}
} // namespace

void convertKVCachePrecisionVector( //TODO: Instead of iterating and calling kernal, concat into a single tensor and call the kernel once
std::vector<runtime::ITensor::SharedPtr>& blocks,
BaseCacheFormatter::CacheState const& srcConfig,
BaseCacheFormatter::CacheState const& destConfig,
runtime::BufferManager const& bufferManager)
{
auto const srcDataType = srcConfig.getDataType();
auto const destDataType = destConfig.getDataType();
auto stream = bufferManager.getStream().get();

TLLM_LOG_INFO("Converting %zu blocks from %s to %s", blocks.size(),
dataTypeToString(srcDataType).c_str(), dataTypeToString(destDataType).c_str());

for (size_t i = 0; i < blocks.size(); ++i)
{
auto& block = blocks[i];
auto blockShape = block->getShape();
auto blockVolume = runtime::ITensor::volume(blockShape);

auto tempBuffer = bufferManager.gpu(blockShape, destDataType);

if (srcDataType == nvinfer1::DataType::kHALF && destDataType == nvinfer1::DataType::kFP8)
{
kernels::invokeConversion<__nv_fp8_e4m3, half>(
reinterpret_cast<__nv_fp8_e4m3*>(tempBuffer->data()),
reinterpret_cast<half const*>(block->data()),
blockVolume,
nullptr,
stream
);
}
else if (srcDataType == nvinfer1::DataType::kBF16 && destDataType == nvinfer1::DataType::kFP8)
{
kernels::invokeConversion<__nv_fp8_e4m3, __nv_bfloat16>(
reinterpret_cast<__nv_fp8_e4m3*>(tempBuffer->data()),
reinterpret_cast<__nv_bfloat16 const*>(block->data()),
blockVolume,
nullptr,
stream
);
}
else if (srcDataType == nvinfer1::DataType::kFP8 && destDataType == nvinfer1::DataType::kHALF)
{
// FP8 -> FP16 conversion for transmission buffer (Overriding for now)
kernels::invokeConversion<half, __nv_fp8_e4m3>(
reinterpret_cast<half*>(tempBuffer->data()),
reinterpret_cast<__nv_fp8_e4m3 const*>(block->data()),
blockVolume,
nullptr,
stream
);
}
else if (srcDataType == nvinfer1::DataType::kBF16 && destDataType == nvinfer1::DataType::kHALF)
{
// BF16 -> FP16 conversion for transmission buffer (Overriding for now)
kernels::invokeConversion<half, __nv_bfloat16>(
reinterpret_cast<half*>(tempBuffer->data()),
reinterpret_cast<__nv_bfloat16 const*>(block->data()),
blockVolume,
nullptr,
stream
);
}
else
{
TLLM_LOG_WARNING("Unsupported conversion %s -> %s, skipping",
dataTypeToString(srcDataType).c_str(),
dataTypeToString(destDataType).c_str());
continue;
}

block = std::move(tempBuffer);
}

bufferManager.getStream().synchronize();
}

void CacheFormatter::unformat(tensorrt_llm::batch_manager::TransferSession& session)
{
NVTX3_SCOPED_RANGE(CacheFormatter_unformat);
Expand Down Expand Up @@ -744,7 +841,14 @@ void CacheFormatter::unformat(tensorrt_llm::batch_manager::TransferSession& sess
{
preAllocRecvBuffer = mCacheTransBufferManager->getRecvBuffer(cacheBufferId);
TLLM_CHECK(preAllocRecvBuffer != nullptr);
TLLM_CHECK(preAllocRecvBuffer->getDataType() == dataType);


// TLLM_CHECK(preAllocRecvBuffer->getDataType() == dataType); <-- KEY ASSERT CHANGED HERE

TLLM_LOG_INFO("============= RECEIVE ON GEN SIDE SETTINGS =============");
TLLM_LOG_INFO("preAllocRecvBuffer data type: %s", dataTypeToString(preAllocRecvBuffer->getDataType()).c_str());
TLLM_LOG_INFO("dataType: %s", dataTypeToString(dataType).c_str());
TLLM_LOG_INFO("============= RECEIVE ON GEN SIDE SETTINGS =============");
}

auto recvBufferFun = [&](int deviceId, size_t processIdx)
Expand Down Expand Up @@ -854,6 +958,15 @@ void CacheFormatter::unformat(tensorrt_llm::batch_manager::TransferSession& sess
{
NVTX3_SCOPED_RANGE(formatInputConcatenate);

if (destConfig.getDataType() != selfConfig.getDataType() && common::getEnvEnableKVCachePrecisionConversion())
{
TLLM_LOG_INFO("WE ARE TAKING THIS PATH, CONVERSING.......");
NVTX3_SCOPED_RANGE(kvCacheRecvPrecisionConv);
convertKVCachePrecisionVector(recvSplitCaches, destConfig, selfConfig, bufferManager);
TLLM_LOG_INFO("After conversion, recvSplitCaches[0] dtype: %s",
dataTypeToString(recvSplitCaches[0]->getDataType()).c_str());
}

executor::kv_cache::concatKvCacheV2Dispatch(
recvSplitCaches, outputBuffersPerWindow, destConfig, selfConfig, selfIdx, bufferManager);

Expand All @@ -873,11 +986,12 @@ void CacheFormatter::unformat(tensorrt_llm::batch_manager::TransferSession& sess

[[nodiscard]] bool CacheFormatter::inquireSupport(CacheState const& selfConfig, CacheState const& destConfig) const
{
if (selfConfig.getDataType() != destConfig.getDataType())
{
TLLM_LOG_WARNING("CacheFormatter::inquireSupport: selfConfig.getDataType() != destConfig.getDataType()");
return false;
}
// TODO: Change
// if (selfConfig.getDataType() != destConfig.getDataType()) // Overriding for now
// {
// TLLM_LOG_WARNING("CacheFormatter::inquireSupport: selfConfig.getDataType() != destConfig.getDataType()");
// return false;
// }

std::unordered_set<SizeType32> setVecSelf{
selfConfig.getModelConfig().mNbKvHeadsPerLayer.begin(), selfConfig.getModelConfig().mNbKvHeadsPerLayer.end()};
Expand Down
97 changes: 78 additions & 19 deletions cpp/tensorrt_llm/batch_manager/cacheTransBuffer.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -203,27 +203,62 @@ CacheTransBufferManager::CacheTransBufferManager(
if (maxNumTokens.has_value())
{
TLLM_CHECK(maxNumTokens.value() % tokensPerBlock == 0);
auto dataSize = common::getDTypeSize(mDataType);
auto kvCacheByteSizePerTokenPerLayer = mCacheManager->getBlockManager().getBlockSize(0) / tokensPerBlock
* (mCacheManager->getCacheType() == CacheType::kSELFKONLY ? 1 : 2) * dataSize;

// Transmission always uses FP16 for mixed precision support
size_t transmissionDataSize = common::getDTypeSize(nvinfer1::DataType::kHALF);
size_t storageDataSize = common::getDTypeSize(mDataType);

TLLM_LOG_INFO("=== Buffer Size Calculation Debug ===");
TLLM_LOG_INFO("maxNumTokens: %ld", maxNumTokens.value());
TLLM_LOG_INFO("tokensPerBlock: %ld", tokensPerBlock);

// getBlockSize() returns volume of [numKvHeads, tokensPerBlock, sizePerHead] for ONE cache (K or V)
// We need to multiply by KV factor (2 for SELF, 1 for SELFKONLY) to get both K and V
size_t blockSizeInElements = mCacheManager->getBlockManager().getBlockSize(0);
size_t kvFactor = (mCacheManager->getCacheType() == CacheType::kSELFKONLY ? 1 : 2);

TLLM_LOG_INFO("getBlockSize(0): %ld elements (for ONE cache)", blockSizeInElements);
TLLM_LOG_INFO("kvFactor: %ld (1=K only, 2=K+V)", kvFactor);
TLLM_LOG_INFO("storageDataSize (mDataType): %ld bytes", storageDataSize);
TLLM_LOG_INFO("transmissionDataSize (FP16): %ld bytes", transmissionDataSize);

// Calculate bytes per token for both K and V in transmission format (FP16)
size_t bytesPerTokenInStorage = (blockSizeInElements * kvFactor * storageDataSize) / tokensPerBlock;
size_t bytesPerTokenInTransmission = (blockSizeInElements * kvFactor * transmissionDataSize) / tokensPerBlock;
auto kvCacheByteSizePerTokenPerLayer = bytesPerTokenInTransmission;

TLLM_LOG_INFO("bytesPerTokenInStorage: %ld", bytesPerTokenInStorage);
TLLM_LOG_INFO("bytesPerTokenInTransmission: %ld", bytesPerTokenInTransmission);
TLLM_LOG_INFO("kvCacheByteSizePerTokenPerLayer: %ld", kvCacheByteSizePerTokenPerLayer);

for (auto layerId = 0; layerId < mCacheManager->getBlockManager().getNumLayers(); layerId++)
{
auto poolIdx = mCacheManager->getBlockManager().getLayerPoolIdx(layerId);
auto windowSize = static_cast<size_t>(mCacheManager->getBlockManager().getPoolWindowSize(poolIdx));
auto alignedWindowSize = (windowSize + tokensPerBlock - 1) / tokensPerBlock * tokensPerBlock;
auto validTokenNum = (alignedWindowSize < maxNumTokens.value() ? alignedWindowSize : maxNumTokens.value());
// if windowSize % (tokensPerBlock) !=0
validTokenNum += tokensPerBlock; // add one more block

auto validTokenNum = (windowSize < maxNumTokens.value() ? windowSize : maxNumTokens.value());

TLLM_LOG_INFO("Layer %d: poolIdx=%d, windowSize=%ld, validTokenNum=%ld",
layerId, poolIdx, windowSize, validTokenNum);
TLLM_LOG_INFO("Layer %d: adding %ld bytes (validTokenNum=%ld * kvCacheBytesPerToken=%ld)",
layerId, validTokenNum * kvCacheByteSizePerTokenPerLayer, validTokenNum, kvCacheByteSizePerTokenPerLayer);

bufferSizeFromMaxNumToken += validTokenNum * kvCacheByteSizePerTokenPerLayer;
}

TLLM_LOG_INFO("Total bufferSizeFromMaxNumToken: %ld bytes", bufferSizeFromMaxNumToken);

}

mTransferBufferSize
= maxNumTokens.has_value() ? bufferSizeFromMaxNumToken : common::getEnvMemSizeForKVCacheTransferBuffer();

TLLM_LOG_INFO("HERE HERE HERE HERE HERE --------- ");
TLLM_LOG_INFO("mTransferBufferSize:%ld", mTransferBufferSize);
TLLM_LOG_INFO("HERE HERE HERE HERE HERE --------- ");

mOnlyUseDynamicBuffer = mTransferBufferSize == 0;
mRecvBufferCount = common::getEnvRequestKVCacheConcurrent() ? common::getEnvKVCacheRecvBufferCount() : 1;
mSendBufferCount = common::getEnvKVCacheSendMaxConcurrenceNum();
mSendBufferCount = common::getEnvParallelCacheSend() ? common::getEnvKVCacheSendMaxConcurrenceNum() : 1;
mUseFabricMemory = !(common::getEnvKVCacheTransferUseSyncBuffer() || common::getEnvKVCacheTransferUseAsyncBuffer())
&& FabricMemory::supportFbaricMemory();
if (mUseFabricMemory)
Expand All @@ -241,6 +276,7 @@ CacheTransBufferManager::CacheTransBufferManager(
allocateBuffer();
}


size_t CacheTransBufferManager::preAllocBufferSize(size_t tokensPerBlock,
std::map<SizeType32, SizeType32> const& cacheSizeBytesPerTokenPerWindow,
std::optional<executor::CacheTransceiverConfig> const& cacheTransceiverConfig)
Expand Down Expand Up @@ -365,8 +401,13 @@ std::tuple<std::vector<runtime::ITensor::SharedPtr>, size_t, bool> CacheTransBuf
}
else
{

retSplitCaches.push_back(bufferManagerToUse.gpu(
runtime::ITensor::makeShape({static_cast<int64_t>(targetBufferEleSizes[i])}), mDataType));
runtime::ITensor::makeShape({static_cast<int64_t>(targetBufferEleSizes[i])}), nvinfer1::DataType::kHALF)); // Overriding for now


// retSplitCaches.push_back(bufferManagerToUse.gpu(
// runtime::ITensor::makeShape({static_cast<int64_t>(targetBufferEleSizes[i])}), mDataType));
}
}
TLLM_LOG_DEBUG("getOrAllocateBuffers bufferCoverTargetNum:%d", bufferCoverTargetNum);
Expand All @@ -384,8 +425,13 @@ std::tuple<std::vector<runtime::ITensor::SharedPtr>, size_t, bool> CacheTransBuf
{
for (int i = 0; i < targetNum; i++)
{

retSplitCaches.push_back(bufferManagerToUse.gpu(
runtime::ITensor::makeShape({static_cast<int64_t>(targetBufferEleSizes[i])}), mDataType));
runtime::ITensor::makeShape({static_cast<int64_t>(targetBufferEleSizes[i])}), nvinfer1::DataType::kHALF));


// retSplitCaches.push_back(bufferManagerToUse.gpu(
// runtime::ITensor::makeShape({static_cast<int64_t>(targetBufferEleSizes[i])}), mDataType)); // Overriding for now
}
bufferCoverTargetNum = targetNum;
}
Expand All @@ -399,36 +445,45 @@ void CacheTransBufferManager::allocateBuffer()
{
return;
}
mBufferEleSize = mTransferBufferSize / common::getDTypeSize(mDataType);
mBufferEleSize = mTransferBufferSize / common::getDTypeSize(nvinfer1::DataType::kHALF); // Overriding for now
// mBufferEleSize = mTransferBufferSize / common::getDTypeSize(mDataType);
mConcurrenceSendResource.mBufferIndexFlag.resize(mSendBufferCount, 0);
mConcurrenceRecvResource.mBufferIndexFlag.resize(mRecvBufferCount, 0);
if (mUseFabricMemory)
{
mFabricMemory.reserve(mSendBufferCount + mRecvBufferCount);
for (size_t i = 0; i < mSendBufferCount; i++)
for (size_t i = 0; i < mSendBufferCount; i++)
{
mFabricMemory.emplace_back(std::make_unique<FabricMemory>(mTransferBufferSize));
mConcurrenceSendResource.mBuffers[i] = runtime::ITensor::wrap(mFabricMemory.back()->getPtr(), mDataType,
mConcurrenceSendResource.mBuffers[i] = runtime::ITensor::wrap(mFabricMemory.back()->getPtr(), nvinfer1::DataType::kHALF, // Overriding for now
runtime::ITensor::makeShape({static_cast<int64_t>(mBufferEleSize)}), mBufferEleSize);
// mConcurrenceSendResource.mBuffers[i] = runtime::ITensor::wrap(mFabricMemory.back()->getPtr(), mDataType,
// runtime::ITensor::makeShape({static_cast<int64_t>(mBufferEleSize)}), mBufferEleSize);
}
for (size_t i = 0; i < mRecvBufferCount; i++)
{
mFabricMemory.emplace_back(std::make_unique<FabricMemory>(mTransferBufferSize));
mConcurrenceRecvResource.mBuffers[i] = runtime::ITensor::wrap(mFabricMemory.back()->getPtr(), mDataType,
mConcurrenceRecvResource.mBuffers[i] = runtime::ITensor::wrap(mFabricMemory.back()->getPtr(), nvinfer1::DataType::kHALF, // Overriding for now
runtime::ITensor::makeShape({static_cast<int64_t>(mBufferEleSize)}), mBufferEleSize);
// mConcurrenceRecvResource.mBuffers[i] = runtime::ITensor::wrap(mFabricMemory.back()->getPtr(), mDataType,
// runtime::ITensor::makeShape({static_cast<int64_t>(mBufferEleSize)}), mBufferEleSize);
}
}
else if (common::getEnvKVCacheTransferUseAsyncBuffer())
{
for (size_t i = 0; i < mSendBufferCount; i++)
{
mConcurrenceSendResource.mBuffers[i]
= mBufferManager.gpu(runtime::ITensor::makeShape({static_cast<int64_t>(mBufferEleSize)}), mDataType);
= mBufferManager.gpu(runtime::ITensor::makeShape({static_cast<int64_t>(mBufferEleSize)}), nvinfer1::DataType::kHALF); // Overriding for now
// mConcurrenceSendResource.mBuffers[i]
// = mBufferManager.gpu(runtime::ITensor::makeShape({static_cast<int64_t>(mBufferEleSize)}), mDataType);
}
for (size_t i = 0; i < mRecvBufferCount; i++)
{
mConcurrenceRecvResource.mBuffers[i]
= mBufferManager.gpu(runtime::ITensor::makeShape({static_cast<int64_t>(mBufferEleSize)}), mDataType);
= mBufferManager.gpu(runtime::ITensor::makeShape({static_cast<int64_t>(mBufferEleSize)}), nvinfer1::DataType::kHALF); // Overriding for now
// mConcurrenceRecvResource.mBuffers[i]
// = mBufferManager.gpu(runtime::ITensor::makeShape({static_cast<int64_t>(mBufferEleSize)}), mDataType);
}
mBufferManager.getStream().synchronize();
}
Expand All @@ -437,12 +492,16 @@ void CacheTransBufferManager::allocateBuffer()
for (size_t i = 0; i < mSendBufferCount; i++)
{
mConcurrenceSendResource.mBuffers[i] = mBufferManager.gpuSync(
runtime::ITensor::makeShape({static_cast<int64_t>(mBufferEleSize)}), mDataType);
runtime::ITensor::makeShape({static_cast<int64_t>(mBufferEleSize)}), nvinfer1::DataType::kHALF); // Overriding for now
// mConcurrenceSendResource.mBuffers[i] = mBufferManager.gpuSync(
// runtime::ITensor::makeShape({static_cast<int64_t>(mBufferEleSize)}), mDataType);
}
for (size_t i = 0; i < mRecvBufferCount; i++)
{
mConcurrenceRecvResource.mBuffers[i] = mBufferManager.gpuSync(
runtime::ITensor::makeShape({static_cast<int64_t>(mBufferEleSize)}), mDataType);
runtime::ITensor::makeShape({static_cast<int64_t>(mBufferEleSize)}), nvinfer1::DataType::kHALF); // Overriding for now
// mConcurrenceRecvResource.mBuffers[i] = mBufferManager.gpuSync(
// runtime::ITensor::makeShape({static_cast<int64_t>(mBufferEleSize)}), mDataType);
}
}
}
Expand Down
6 changes: 6 additions & 0 deletions cpp/tensorrt_llm/common/envUtils.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -450,4 +450,10 @@ bool getEnvDisableChunkedAttentionInGenPhase()
return getBoolEnv("TRTLLM_DISABLE_CHUNKED_ATTENTION_IN_GEN_PHASE");
}

bool getEnvEnableKVCachePrecisionConversion()
{
static bool const enableKVCachePrecisionConversion = getBoolEnv("TRTLLM_ENABLE_KVCACHE_PRECISION_CONVERSION");
return enableKVCachePrecisionConversion;
}

} // namespace tensorrt_llm::common
3 changes: 3 additions & 0 deletions cpp/tensorrt_llm/common/envUtils.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -136,4 +136,7 @@ bool getEnvDisaggBenchmarkGenOnly();
// Whether to disable the chunked-attention in the generation phase.
bool getEnvDisableChunkedAttentionInGenPhase();

// Whether to enable KV cache precision conversion during transfers.
bool getEnvEnableKVCachePrecisionConversion();

} // namespace tensorrt_llm::common
Loading
Loading