mirror of
https://github.com/NVIDIA/TensorRT-LLM.git
synced 2026-01-14 06:27:45 +08:00
b7a38feb14
* add space in test output Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com> * perf: reduce executor lock scope Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com> * refactor: Move TokenRangeRetentionConfig implementation to cpp file Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com> * fix: Improve finished steps handling for external draft tokens - Fixed a bug where the whole finished steps tensor was being zeroes instead of the slices. - Replaced the creation of a temporary tensor for finished steps with a direct slice from the input tensor, improving efficiency and readability. - Updated the tensor management logic to streamline the process of setting zero values for finished steps during batch processing. Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com> * chore: Clean up includes Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com> --------- Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com>
68 lines
2.6 KiB
C++
68 lines
2.6 KiB
C++
/*
|
|
* Copyright (c) 2022-2024, 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/runtime/iStatefulGptDecoder.h"
|
|
|
|
namespace tensorrt_llm::runtime
|
|
{
|
|
|
|
class GptDecoderBatched;
|
|
|
|
//! GPT decoder class with support for in-flight batching and stateful interface
|
|
class StatefulGptDecoderBatched : public IStatefulGptDecoder
|
|
{
|
|
public:
|
|
using CudaStreamPtr = std::shared_ptr<CudaStream>;
|
|
using TensorPtr = ITensor::SharedPtr;
|
|
|
|
StatefulGptDecoderBatched(CudaStreamPtr stream, nvinfer1::DataType dtype);
|
|
|
|
~StatefulGptDecoderBatched() override;
|
|
|
|
// IStatefulGptDecoder implementation
|
|
void setup(executor::DecodingMode const& mode, SizeType32 maxBatchSize, SizeType32 maxBeamWidth,
|
|
SizeType32 maxAttentionWindow, SizeType32 sinkTokenLength, SizeType32 maxSequenceLength,
|
|
nvinfer1::DataType dtype, ModelConfig const& modelConfig, WorldConfig const& worldConfig) override;
|
|
|
|
void newBatch(GenerationInput const& inputs, GenerationOutput const& outputs, SamplingConfig const& samplingConfig,
|
|
ModelConfig const& modelConfig) override;
|
|
|
|
void forwardAsync(decoder::Output& output, decoder::Input const& input) override;
|
|
void forwardSync() override;
|
|
|
|
[[nodiscard]] TensorPtr getIds() const override;
|
|
[[nodiscard]] TensorPtr getGatheredIds() const override;
|
|
void finalize(SamplingConfig const& samplingConfig) const override;
|
|
[[nodiscard]] TensorPtr getCumLogProbs() const override;
|
|
[[nodiscard]] TensorPtr getLogProbs() const override;
|
|
[[nodiscard]] TensorPtr getNewTokens(SizeType32 iter = 0) const override;
|
|
[[nodiscard]] TensorPtr getNbFinished() const override;
|
|
|
|
private:
|
|
std::unique_ptr<GptDecoderBatched> mDecoder;
|
|
|
|
// only used for IStatefulGptDecoder
|
|
CudaEvent mDecoderFinishEvent;
|
|
CudaEvent mForwardEvent;
|
|
TensorPtr mFinishedSum;
|
|
TensorPtr mBatchSlotsSetup; // [maxBatchSize], int32_t, address map, pinned
|
|
TensorPtr mBatchSlotsDecoder; // [maxBatchSize], int32_t, address map, pinned
|
|
};
|
|
|
|
} // namespace tensorrt_llm::runtime
|