mirror of
https://github.com/NVIDIA/TensorRT-LLM.git
synced 2026-01-14 06:27:45 +08:00
114 lines
4.1 KiB
C++
114 lines
4.1 KiB
C++
/*
|
|
* SPDX-FileCopyrightText: Copyright (c) 2025 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 "tensorrt_llm/executor/dynamicBatchTuner.h"
|
|
#include "tensorrt_llm/common/logger.h"
|
|
#include <cmath>
|
|
|
|
namespace
|
|
{
|
|
using namespace tensorrt_llm::executor;
|
|
|
|
void updateStats(SizeType32 value, std::deque<SizeType32>& stats, int64_t& sum, SizeType32 windowSize)
|
|
{
|
|
while (static_cast<SizeType32>(stats.size()) >= windowSize)
|
|
{
|
|
sum -= stats.front();
|
|
stats.pop_front();
|
|
}
|
|
stats.push_back(value);
|
|
sum += value;
|
|
}
|
|
} // namespace
|
|
|
|
namespace tensorrt_llm::executor
|
|
{
|
|
|
|
DynamicBatchTuner::DynamicBatchTuner(DynamicBatchConfig const& config)
|
|
: mEnableBatchSizeTuning(config.getEnableBatchSizeTuning())
|
|
, mEnableMaxNumTokensTuning(config.getEnableMaxNumTokensTuning())
|
|
, mDynamicBatchMovingAverageWindow(config.getDynamicBatchMovingAverageWindow())
|
|
, mBatchSizeTable(config.getBatchSizeTable())
|
|
{
|
|
TLLM_CHECK_WITH_INFO(!mBatchSizeTable.empty(), "Batch size table is empty.");
|
|
for (size_t i = 1; i < mBatchSizeTable.size(); ++i)
|
|
{
|
|
TLLM_CHECK_WITH_INFO(mBatchSizeTable[i - 1].first < mBatchSizeTable[i].first,
|
|
"Batch size table is not sorted in ascending order.");
|
|
}
|
|
}
|
|
|
|
void DynamicBatchTuner::updateStats(SizeType32 inputLength, SizeType32 outputLength)
|
|
{
|
|
::updateStats(inputLength, mInputLengthStats, mInputLengthSum, mDynamicBatchMovingAverageWindow);
|
|
::updateStats(outputLength, mOutputLengthStats, mOutputLengthSum, mDynamicBatchMovingAverageWindow);
|
|
}
|
|
|
|
double DynamicBatchTuner::getAverageInputLength() const
|
|
{
|
|
return mInputLengthStats.empty() ? 0 : static_cast<double>(mInputLengthSum) / mInputLengthStats.size();
|
|
}
|
|
|
|
double DynamicBatchTuner::getAverageOutputLength() const
|
|
{
|
|
return mOutputLengthStats.empty() ? 0 : static_cast<double>(mOutputLengthSum) / mOutputLengthStats.size();
|
|
}
|
|
|
|
SizeType32 DynamicBatchTuner::getRuntimeBatchSize(SizeType32 maxCapacityBatchSize) const
|
|
{
|
|
for (auto const& [batchSizeLimit, batchSize] : mBatchSizeTable)
|
|
{
|
|
if (maxCapacityBatchSize < batchSizeLimit)
|
|
{
|
|
return batchSize;
|
|
}
|
|
}
|
|
SizeType32 threshold = maxCapacityBatchSize / kBatchSizeFallbackGranularity * kBatchSizeFallbackGranularity;
|
|
if (maxCapacityBatchSize < (threshold + kBatchSizeFallbackThreshold))
|
|
{
|
|
return threshold;
|
|
}
|
|
return maxCapacityBatchSize;
|
|
}
|
|
|
|
SizeType32 DynamicBatchTuner::getRuntimeMaxNumTokens(SizeType32 maxRuntimeBatchSize) const
|
|
{
|
|
// calculate max num token in fully overlapped case
|
|
SizeType32 adjustedNumTokens
|
|
= 1.0 * (maxRuntimeBatchSize * getAverageInputLength() / getAverageOutputLength() + maxRuntimeBatchSize);
|
|
SizeType32 tokenThreshold;
|
|
// context heavy (avg ISL/OSL > kMaxNumTokensRatioContextHeavy)
|
|
if (getAverageInputLength() / getAverageOutputLength() > kMaxNumTokensRatioContextHeavy)
|
|
{
|
|
tokenThreshold = kMaxNumTokensThresholdContextHeavy;
|
|
}
|
|
// balanced case (kMaxNumTokensRatioBalanced < avg ISL/OSL < kMaxNumTokensRatioContextHeavy)
|
|
else if (getAverageInputLength() / getAverageOutputLength() > kMaxNumTokensRatioBalanced)
|
|
{
|
|
tokenThreshold = kMaxNumTokensThresholdBalanced;
|
|
}
|
|
// gen heavy (avg ISL/OSL < kMaxNumTokensRatioBalanced)
|
|
else
|
|
{
|
|
tokenThreshold = kMaxNumTokensThresholdGenHeavy;
|
|
}
|
|
// pad it to pow of 2 and max of this value and threshold.
|
|
return (std::max(1 << int(ceil(log2(adjustedNumTokens))), tokenThreshold));
|
|
}
|
|
|
|
} // namespace tensorrt_llm::executor
|