mirror of
https://github.com/NVIDIA/TensorRT-LLM.git
synced 2026-01-22 19:52:38 +08:00
250d9c293d
* Update TensorRT-LLM --------- Co-authored-by: Bhuvanesh Sridharan <bhuvan.sridharan@gmail.com> Co-authored-by: Morgan Funtowicz <funtowiczmo@gmail.com> Co-authored-by: Eddie-Wang1120 <wangjinheng1120@163.com> Co-authored-by: meghagarwal <16129366+megha95@users.noreply.github.com>
96 lines
3.0 KiB
C++
96 lines
3.0 KiB
C++
/*
|
|
* Copyright (c) 2019-2024, NVIDIA CORPORATION. All rights reserved.
|
|
* Copyright (c) 2021, NAVER Corp. Authored by CLOVA.
|
|
*
|
|
* 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 <curand_kernel.h>
|
|
|
|
#include "tensorrt_llm/common/tensor.h"
|
|
#include "tensorrt_llm/layers/baseSamplingLayer.h"
|
|
#include "tensorrt_llm/layers/decodingParams.h"
|
|
#include "tensorrt_llm/layers/topKSamplingLayer.h"
|
|
#include "tensorrt_llm/layers/topPSamplingLayer.h"
|
|
#include "tensorrt_llm/runtime/decodingMode.h"
|
|
|
|
namespace tc = tensorrt_llm::common;
|
|
|
|
namespace tensorrt_llm
|
|
{
|
|
namespace layers
|
|
{
|
|
|
|
template <typename T>
|
|
inline bool allOfBatchSlots(
|
|
runtime::SizeType const* batchSlotsHost, T const* data, runtime::SizeType batchSize, T value)
|
|
{
|
|
return std::all_of(
|
|
batchSlotsHost, batchSlotsHost + batchSize, [&](runtime::SizeType b) { return data[b] == value; });
|
|
};
|
|
|
|
//! \brief Top class for sampling layers.
|
|
//! It sets up and executes TopKSamplingLayer and TopPSamplingLayer samplings
|
|
template <typename T>
|
|
class SamplingLayer : public BaseSamplingLayer<T>
|
|
{
|
|
public:
|
|
using Base = BaseSamplingLayer<T>;
|
|
using SetupParams = typename Base::SetupParams;
|
|
using ForwardParams = typename Base::ForwardParams;
|
|
|
|
SamplingLayer(runtime::DecodingMode const& mode, runtime::SizeType maxBatchSize, runtime::SizeType vocabSize,
|
|
runtime::SizeType vocabSizePadded, cudaStream_t stream,
|
|
std::shared_ptr<tensorrt_llm::common::IAllocator> allocator, cudaDeviceProp* prop);
|
|
|
|
~SamplingLayer() override = default;
|
|
|
|
void forward(DecodingOutputParams& outputs, ForwardParams& inputs) override;
|
|
|
|
void setup(
|
|
runtime::SizeType batchSize, runtime::SizeType const* batchSlots, SetupParams const& setupParams) override;
|
|
|
|
private:
|
|
using Base::mMaxBatchSize;
|
|
using Base::mVocabSize;
|
|
using Base::mVocabSizePadded;
|
|
using Base::mSamplingWorkspaceSize;
|
|
using Base::mAllocatedSize;
|
|
|
|
using Base::mStream;
|
|
using Base::mAllocator;
|
|
|
|
runtime::DecodingMode mDecodingMode;
|
|
|
|
void* mSamplingWorkspaceDevice = nullptr;
|
|
curandState_t* mCurandStatesDevice = nullptr;
|
|
uint64_t* mRandomSeedsDevice = nullptr;
|
|
|
|
bool* mSkipDecodeDevice = nullptr;
|
|
|
|
bool* mSkipDecodeHost = nullptr;
|
|
bool mSkipAny = false;
|
|
|
|
std::unique_ptr<TopKSamplingLayer<T>> mTopKDecode;
|
|
std::unique_ptr<TopPSamplingLayer<T>> mTopPDecode;
|
|
|
|
private:
|
|
void allocateBuffer(runtime::SizeType batchSize);
|
|
void freeBuffer();
|
|
};
|
|
|
|
} // namespace layers
|
|
} // namespace tensorrt_llm
|