mirror of
https://github.com/NVIDIA/TensorRT-LLM.git
synced 2026-01-14 06:27:45 +08:00
f430a4b447
* Update TensorRT-LLM --------- Co-authored-by: IbrahimAmin <ibrahimamin532@gmail.com> Co-authored-by: Fabian Joswig <fjosw@users.noreply.github.com> Co-authored-by: Pzzzzz <hello-cd.plus@hotmail.com> Co-authored-by: CoderHam <hemant@cohere.com> Co-authored-by: Konstantin Lopuhin <kostia.lopuhin@gmail.com>
73 lines
2.8 KiB
C++
73 lines
2.8 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/executor/types.h"
|
|
#include "tensorrt_llm/layers/baseLayer.h"
|
|
#include "tensorrt_llm/layers/decodingParams.h"
|
|
#include "tensorrt_llm/runtime/iTensor.h"
|
|
|
|
namespace tensorrt_llm
|
|
{
|
|
namespace layers
|
|
{
|
|
|
|
//! \brief Layer to ban specific words from being sampled.
|
|
//! Supports banning bad words and repeating N grams.
|
|
//! Set badWordsPtr, maxBadWordsLen and badWordsLengths to ban bad words.
|
|
//! Set noRepeatNgramSize in input params to ban repeat Ngrams.
|
|
//! Layer modifies logits in-place.
|
|
template <typename T>
|
|
class BanWordsLayer : public BaseLayer
|
|
{
|
|
public:
|
|
BanWordsLayer(executor::DecodingMode const& mode, DecoderDomain const& decoderDomain, cudaStream_t stream,
|
|
std::shared_ptr<tensorrt_llm::common::IAllocator> allocator);
|
|
|
|
~BanWordsLayer() override = default;
|
|
|
|
void setup(runtime::SizeType32 batchSize, runtime::SizeType32 beamWidth, runtime::SizeType32 const* batchSlots,
|
|
std::shared_ptr<BaseSetupParams> setupParams) override;
|
|
|
|
//! \brief Modifies 'outputs->logits' in-place with -INF for banned words
|
|
void forwardAsync(std::shared_ptr<BaseOutputParams> outputs, std::shared_ptr<BaseInputParams> inputs) override;
|
|
|
|
private:
|
|
static void banRepeatNGrams(tc::Tensor& logits, std::shared_ptr<DynamicDecodeOutputParams> const& outputs,
|
|
std::shared_ptr<DynamicDecodeInputParams> const& params, runtime::SizeType32 const* batchSlots,
|
|
DecoderDomain const& decoderDomain, runtime::SizeType32 maxSeqLen, cudaStream_t stream);
|
|
static void banBadWords(tc::Tensor& logits, std::shared_ptr<DynamicDecodeOutputParams> const& outputs,
|
|
std::shared_ptr<DynamicDecodeInputParams> const& params, runtime::SizeType32 const* batchSlots,
|
|
DecoderDomain const& decoderDomain, runtime::SizeType32 maxSeqLen, cudaStream_t stream);
|
|
|
|
private:
|
|
using BaseLayer::mWorkspaceSize;
|
|
using BaseLayer::mAllocatedSize;
|
|
|
|
using BaseLayer::mStream;
|
|
using BaseLayer::mAllocator;
|
|
|
|
executor::DecodingMode mDecodingMode;
|
|
};
|
|
|
|
} // namespace layers
|
|
} // namespace tensorrt_llm
|