mirror of
https://github.com/NVIDIA/TensorRT-LLM.git
synced 2026-01-14 06:27:45 +08:00
393 lines
15 KiB
C++
393 lines
15 KiB
C++
/*
|
|
* Copyright (c) 2022-2023, 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.
|
|
*/
|
|
#ifndef TOP_LEVEL_DIR
|
|
#error "Define TOP_LEVEL_DIR"
|
|
#endif
|
|
|
|
#include <gtest/gtest.h>
|
|
|
|
#include "tensorrt_llm/common/memoryUtils.h"
|
|
#include "tensorrt_llm/kernels/decodingCommon.h"
|
|
#include "tensorrt_llm/kernels/stopCriteriaKernels.h"
|
|
#include "tensorrt_llm/runtime/bufferManager.h"
|
|
#include <algorithm>
|
|
#include <curand_kernel.h>
|
|
#include <random>
|
|
|
|
namespace tk = tensorrt_llm::kernels;
|
|
namespace tc = tensorrt_llm::common;
|
|
|
|
using namespace tensorrt_llm::runtime;
|
|
|
|
namespace
|
|
{
|
|
|
|
class StopCriteriaKernelsTest : public testing::Test
|
|
{
|
|
public:
|
|
using TensorPtr = tensorrt_llm::runtime::ITensor::SharedPtr;
|
|
|
|
void SetUp() override
|
|
{
|
|
mStream = std::make_shared<tensorrt_llm::runtime::CudaStream>();
|
|
mBufferManager = std::make_shared<tensorrt_llm::runtime::BufferManager>(mStream);
|
|
}
|
|
|
|
void TearDown() override {}
|
|
|
|
void initData(SizeType seed, const std::vector<std::vector<std::vector<SizeType>>>& stopWords,
|
|
SizeType stopWordsLen, SizeType batchSize, SizeType beamWidth)
|
|
{
|
|
std::mt19937 generator(seed);
|
|
std::uniform_int_distribution<int> seqLenDistr(0, mMaxSeqLen);
|
|
|
|
mSequenceLengths
|
|
= mBufferManager->pinned(ITensor::makeShape({batchSize, beamWidth}), nvinfer1::DataType::kINT32);
|
|
mSequenceLengthLimits = mBufferManager->pinned(ITensor::makeShape({batchSize}), nvinfer1::DataType::kINT32);
|
|
mFinished = mBufferManager->pinned(
|
|
ITensor::makeShape({batchSize, beamWidth}), TRTDataType<tk::FinishedState::UnderlyingType>::value);
|
|
mFinishedSum = mBufferManager->pinned(ITensor::makeShape({1}), nvinfer1::DataType::kINT32);
|
|
|
|
mOutputIds = mBufferManager->pinned(
|
|
ITensor::makeShape({batchSize, beamWidth, mMaxSeqLen}), nvinfer1::DataType::kINT32);
|
|
mOutputIdsPtr = mBufferManager->pinned(ITensor::makeShape({batchSize, beamWidth}), nvinfer1::DataType::kINT64);
|
|
|
|
mParentIds = mBufferManager->pinned(
|
|
ITensor::makeShape({batchSize, beamWidth, mMaxSeqLen}), nvinfer1::DataType::kINT32);
|
|
mParentIdsPtr = mBufferManager->pinned(ITensor::makeShape({batchSize, beamWidth}), nvinfer1::DataType::kINT64);
|
|
|
|
mRefOutputIds = mBufferManager->pinned(
|
|
ITensor::makeShape({batchSize, beamWidth, mMaxSeqLen}), nvinfer1::DataType::kINT32);
|
|
|
|
mStopWords
|
|
= mBufferManager->pinned(ITensor::makeShape({batchSize, 2, stopWordsLen}), nvinfer1::DataType::kINT32);
|
|
|
|
auto sequenceLengthsPtr = bufferCast<SizeType>(*mSequenceLengths);
|
|
auto sequenceLengthLimitsPtr = bufferCast<SizeType>(*mSequenceLengthLimits);
|
|
auto finishedPtr
|
|
= reinterpret_cast<tk::FinishedState*>(bufferCast<tk::FinishedState::UnderlyingType>(*mFinished));
|
|
auto finishedSumPtr = bufferCast<SizeType>(*mFinishedSum);
|
|
|
|
for (SizeType bi = 0; bi < batchSize; ++bi)
|
|
{
|
|
for (SizeType ri = 0; ri < beamWidth; ri++)
|
|
{
|
|
sequenceLengthsPtr[bi * beamWidth + ri]
|
|
= stopWordsLen == 0 ? seqLenDistr(generator) : mMaxSeqLen - (bi + ri) % mMaxSeqLen;
|
|
finishedPtr[bi * beamWidth + ri] = tk::FinishedState::empty();
|
|
}
|
|
}
|
|
for (SizeType bi = 0; bi < batchSize; ++bi)
|
|
{
|
|
sequenceLengthLimitsPtr[bi] = stopWordsLen == 0 ? seqLenDistr(generator) : mMaxSeqLen - bi % mMaxSeqLen;
|
|
}
|
|
finishedSumPtr[0] = 0;
|
|
|
|
auto outputIdsPtrsData = reinterpret_cast<void**>(bufferCast<int64_t>(*mOutputIdsPtr));
|
|
auto parentIdsPtrsData = reinterpret_cast<void**>(bufferCast<int64_t>(*mParentIdsPtr));
|
|
auto outputIdsData = bufferCast<int32_t>(*mOutputIds);
|
|
auto refOutputIdsData = bufferCast<int32_t>(*mRefOutputIds);
|
|
auto parentIdsData = bufferCast<int32_t>(*mParentIds);
|
|
|
|
// Tokens ids are
|
|
// bi: 0, ri: 0: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]
|
|
// bi: 0, ri: 1: [16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30]
|
|
// bi: 1, ri: 0: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14]
|
|
// bi: 1, ri: 1: [16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29]
|
|
// bi: 2, ri: 0: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13]
|
|
// bi: 2, ri: 1: [16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28]
|
|
// bi: 3, ri: 0: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12]
|
|
// bi: 3, ri: 1: [16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27]
|
|
for (SizeType bi = 0; bi < batchSize; bi++)
|
|
{
|
|
for (SizeType ri = 0; ri < beamWidth; ri++)
|
|
{
|
|
for (SizeType si = 0; si < mMaxSeqLen; si++)
|
|
{
|
|
auto const idx = tc::flat_index3(bi, ri, si, beamWidth, mMaxSeqLen);
|
|
outputIdsData[idx] = ri * mMaxSeqLen + si;
|
|
parentIdsData[idx] = 0;
|
|
}
|
|
}
|
|
}
|
|
|
|
for (SizeType bi = 0; bi < batchSize; bi++)
|
|
{
|
|
outputIdsPtrsData[bi] = outputIdsData + bi * beamWidth * mMaxSeqLen;
|
|
parentIdsPtrsData[bi] = parentIdsData + bi * beamWidth * mMaxSeqLen;
|
|
}
|
|
|
|
// Init stop words tensor
|
|
auto stopWordsData = bufferCast<int32_t>(*mStopWords);
|
|
std::fill(stopWordsData, stopWordsData + batchSize * 2 * stopWordsLen, -1);
|
|
for (SizeType bi = 0; bi < stopWords.size(); bi++)
|
|
{
|
|
SizeType totalLen = 0;
|
|
for (SizeType wi = 0; wi < stopWords[bi].size(); ++wi)
|
|
{
|
|
for (SizeType si = 0; si < stopWords[bi][wi].size(); ++si)
|
|
{
|
|
stopWordsData[bi * 2 * stopWordsLen + 0 * stopWordsLen + totalLen + si] = stopWords[bi][wi][si];
|
|
}
|
|
totalLen += stopWords[bi][wi].size();
|
|
// Do not add value if stop words is empty
|
|
if (totalLen > 0)
|
|
{
|
|
stopWordsData[bi * 2 * stopWordsLen + 1 * stopWordsLen + wi] = totalLen;
|
|
}
|
|
}
|
|
// Special case when all stop words are of single token length
|
|
if (stopWords[bi].size() == totalLen)
|
|
{
|
|
stopWordsData[bi * 2 * stopWordsLen + 1 * stopWordsLen + totalLen] = totalLen + 1;
|
|
}
|
|
}
|
|
}
|
|
|
|
void verifyMaxSeqLenStopCriteriaResults(SizeType seed, SizeType batchSize, SizeType beamWidth)
|
|
{
|
|
mStream->synchronize();
|
|
|
|
auto sequenceLengthsPtr = bufferCast<SizeType>(*mSequenceLengths);
|
|
auto sequenceLengthLimitsPtr = bufferCast<SizeType>(*mSequenceLengthLimits);
|
|
auto finishedPtr
|
|
= reinterpret_cast<tk::FinishedState*>(bufferCast<tk::FinishedState::UnderlyingType>(*mFinished));
|
|
auto finishedSumPtr = bufferCast<SizeType>(*mFinishedSum);
|
|
|
|
int32_t refSumFinished = 0;
|
|
for (SizeType bi = 0; bi < batchSize * beamWidth; ++bi)
|
|
{
|
|
auto const batchIdx = bi / beamWidth;
|
|
auto const beamIdx = bi % beamWidth;
|
|
const auto limitExceeded = sequenceLengthsPtr[bi] >= sequenceLengthLimitsPtr[batchIdx];
|
|
refSumFinished += limitExceeded;
|
|
if (limitExceeded)
|
|
{
|
|
EXPECT_TRUE(finishedPtr[bi].isFinishedMaxLength())
|
|
<< " batchIdx: " << batchIdx << " beamIdx: " << beamIdx << " seed: " << seed;
|
|
}
|
|
}
|
|
EXPECT_EQ(refSumFinished, finishedSumPtr[0]);
|
|
}
|
|
|
|
bool isSubsequence(const SizeType* sequence, SizeType n, const std::vector<int>& subsequence)
|
|
{
|
|
auto it = std::search(sequence, sequence + n, subsequence.begin(), subsequence.end());
|
|
return it != sequence + n;
|
|
}
|
|
|
|
void verifyStopWordsStopCriteriaResults(SizeType seed,
|
|
const std::vector<std::vector<std::vector<SizeType>>>& stopWords, SizeType stopWordsLen, SizeType batchSize,
|
|
SizeType beamWidth)
|
|
{
|
|
mStream->synchronize();
|
|
|
|
auto outputIdsData = bufferCast<int32_t>(*mOutputIds);
|
|
auto finishedPtr
|
|
= reinterpret_cast<tk::FinishedState*>(bufferCast<tk::FinishedState::UnderlyingType>(*mFinished));
|
|
auto sequenceLengthsPtr = bufferCast<SizeType>(*mSequenceLengths);
|
|
|
|
for (SizeType bi = 0; bi < batchSize; bi++)
|
|
{
|
|
for (SizeType bwi = 0; bwi < beamWidth; bwi++)
|
|
{
|
|
auto outputIdsBatchBeam = outputIdsData + bi * beamWidth * mMaxSeqLen + bwi * mMaxSeqLen;
|
|
bool found = false;
|
|
for (SizeType wi = 0; wi < stopWords[bi].size(); ++wi)
|
|
{
|
|
auto const wordLen = stopWords[bi][wi].size();
|
|
auto const seqLen = sequenceLengthsPtr[bi * beamWidth + bwi];
|
|
auto const offset = seqLen - wordLen;
|
|
found |= isSubsequence(outputIdsBatchBeam + offset, wordLen, stopWords[bi][wi]);
|
|
if (found)
|
|
{
|
|
EXPECT_TRUE(finishedPtr[bi * beamWidth + bwi].isFinishedStopWords());
|
|
break;
|
|
}
|
|
}
|
|
if (!found)
|
|
{
|
|
EXPECT_FALSE(finishedPtr[bi * beamWidth + bwi].isFinished());
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
void runStopWordsCriteriaTest(
|
|
const std::vector<std::vector<std::vector<SizeType>>>& stopWords, SizeType batchSize, SizeType beamWidth)
|
|
{
|
|
SizeType maxStopWordsLen = 0;
|
|
for (const auto& batchStopWords : stopWords)
|
|
{
|
|
SizeType stopWordsLen = 0;
|
|
for (const auto& words : batchStopWords)
|
|
{
|
|
stopWordsLen += words.size();
|
|
}
|
|
if (stopWordsLen == batchStopWords.size())
|
|
{
|
|
stopWordsLen += 1;
|
|
}
|
|
maxStopWordsLen = std::max(maxStopWordsLen, stopWordsLen);
|
|
}
|
|
|
|
initData(0, stopWords, maxStopWordsLen, batchSize, beamWidth);
|
|
|
|
tk::invokeStopWordsCriterion(reinterpret_cast<const int**>(bufferCast<int64_t>(*mOutputIdsPtr)),
|
|
reinterpret_cast<const int**>(bufferCast<int64_t>(*mParentIdsPtr)), bufferCast<SizeType>(*mStopWords),
|
|
reinterpret_cast<tk::FinishedState*>(bufferCast<tk::FinishedState::UnderlyingType>(*mFinished)),
|
|
bufferCast<SizeType>(*mSequenceLengths), maxStopWordsLen, batchSize, beamWidth, mMaxSeqLen, mStream->get());
|
|
|
|
verifyStopWordsStopCriteriaResults(0, stopWords, maxStopWordsLen, batchSize, beamWidth);
|
|
}
|
|
|
|
void runMaxLengthCriteriaTest(SizeType seed, SizeType batchSize, SizeType beamWidth)
|
|
{
|
|
initData(seed, {}, 0, batchSize, beamWidth);
|
|
|
|
tk::invokeLengthCriterion(
|
|
reinterpret_cast<tk::FinishedState*>(bufferCast<tk::FinishedState::UnderlyingType>(*mFinished)),
|
|
bufferCast<SizeType>(*mFinishedSum),
|
|
reinterpret_cast<const uint32_t*>(bufferCast<SizeType>(*mSequenceLengthLimits)),
|
|
bufferCast<SizeType>(*mSequenceLengths), batchSize, beamWidth, mStream->get());
|
|
|
|
verifyMaxSeqLenStopCriteriaResults(seed, batchSize, beamWidth);
|
|
}
|
|
|
|
protected:
|
|
std::shared_ptr<tensorrt_llm::runtime::BufferManager> mBufferManager;
|
|
std::shared_ptr<tensorrt_llm::runtime::CudaStream> mStream;
|
|
|
|
TensorPtr mSequenceLengths;
|
|
TensorPtr mSequenceLengthLimits;
|
|
TensorPtr mFinished;
|
|
TensorPtr mFinishedSum;
|
|
|
|
TensorPtr mOutputIds;
|
|
TensorPtr mRefOutputIds;
|
|
TensorPtr mOutputIdsPtr;
|
|
TensorPtr mParentIds;
|
|
TensorPtr mParentIdsPtr;
|
|
TensorPtr mStopWords;
|
|
|
|
static constexpr SizeType mMaxSeqLen{16};
|
|
static constexpr SizeType mVocabSize{32};
|
|
};
|
|
|
|
TEST_F(StopCriteriaKernelsTest, maxLengthCriteriaBS1BW1Test)
|
|
{
|
|
constexpr SizeType seeds = 64;
|
|
constexpr SizeType batchSize = 1;
|
|
constexpr SizeType beamWidth = 1;
|
|
for (SizeType seed = 0; seed < seeds; ++seed)
|
|
{
|
|
this->runMaxLengthCriteriaTest(seed, batchSize, beamWidth);
|
|
}
|
|
}
|
|
|
|
TEST_F(StopCriteriaKernelsTest, maxLengthCriteriaBS1BW2Test)
|
|
{
|
|
constexpr SizeType seeds = 64;
|
|
constexpr SizeType batchSize = 1;
|
|
constexpr SizeType beamWidth = 2;
|
|
for (SizeType seed = 0; seed < seeds; ++seed)
|
|
{
|
|
this->runMaxLengthCriteriaTest(seed, batchSize, beamWidth);
|
|
}
|
|
}
|
|
|
|
TEST_F(StopCriteriaKernelsTest, maxLengthCriteriaBS1024BW1Test)
|
|
{
|
|
constexpr SizeType seeds = 64;
|
|
constexpr SizeType batchSize = 1024;
|
|
constexpr SizeType beamWidth = 1;
|
|
for (SizeType seed = 0; seed < seeds; ++seed)
|
|
{
|
|
this->runMaxLengthCriteriaTest(seed, batchSize, beamWidth);
|
|
}
|
|
}
|
|
|
|
TEST_F(StopCriteriaKernelsTest, maxLengthCriteriaBS1024BW2Test)
|
|
{
|
|
constexpr SizeType seeds = 64;
|
|
constexpr SizeType batchSize = 1024;
|
|
constexpr SizeType beamWidth = 2;
|
|
for (SizeType seed = 0; seed < seeds; ++seed)
|
|
{
|
|
this->runMaxLengthCriteriaTest(seed, batchSize, beamWidth);
|
|
}
|
|
}
|
|
|
|
TEST_F(StopCriteriaKernelsTest, stopWordsCriteriaBS1SingleTokenSingleWordTest)
|
|
{
|
|
constexpr SizeType batchSize = 1;
|
|
constexpr SizeType beamWidth = 1;
|
|
// Expected to not match any word
|
|
this->runStopWordsCriteriaTest({{{2}}}, batchSize, beamWidth);
|
|
}
|
|
|
|
TEST_F(StopCriteriaKernelsTest, stopWordsCriteriaBS1SingleTokenMultipleWordsTest)
|
|
{
|
|
constexpr SizeType batchSize = 1;
|
|
constexpr SizeType beamWidth = 1;
|
|
// Expected to match 15
|
|
this->runStopWordsCriteriaTest({{{145}, {4}, {1}, {15}}}, batchSize, beamWidth);
|
|
}
|
|
|
|
TEST_F(StopCriteriaKernelsTest, stopWordsCriteriaBS1MultipleTokensSingleWordTest)
|
|
{
|
|
constexpr SizeType batchSize = 1;
|
|
constexpr SizeType beamWidth = 1;
|
|
// Expected to not match any word
|
|
this->runStopWordsCriteriaTest({{{2, 3}}}, batchSize, beamWidth);
|
|
}
|
|
|
|
TEST_F(StopCriteriaKernelsTest, stopWordsCriteriaBS1MultipleTokensMultipleWordsMatchTest)
|
|
{
|
|
constexpr SizeType batchSize = 1;
|
|
constexpr SizeType beamWidth = 1;
|
|
// Expected to match {13, 14, 15}
|
|
this->runStopWordsCriteriaTest({{{1, 4}, {2, 3}, {13, 14, 15}}}, batchSize, beamWidth);
|
|
}
|
|
|
|
TEST_F(StopCriteriaKernelsTest, stopWordsCriteriaBS1MultipleTokensMultipleWordsNotMatchTest)
|
|
{
|
|
constexpr SizeType batchSize = 1;
|
|
constexpr SizeType beamWidth = 1;
|
|
// Expected to not match any word
|
|
this->runStopWordsCriteriaTest({{{1, 4}, {2, 3}, {12, 14, 15}}}, batchSize, beamWidth);
|
|
}
|
|
|
|
TEST_F(StopCriteriaKernelsTest, stopWordsCriteriaBS4MultipleTokensMultipleWordsTest)
|
|
{
|
|
constexpr SizeType batchSize = 4;
|
|
constexpr SizeType beamWidth = 1;
|
|
// Expected to match {12, 13} for the 3rd instance in the batch
|
|
this->runStopWordsCriteriaTest({{{2}}, {{}}, {{15}, {12, 13}}, {{1}, {8, 9}}}, batchSize, beamWidth);
|
|
}
|
|
|
|
TEST_F(StopCriteriaKernelsTest, stopWordsCriteriaBS4BW2MultipleTokensMultipleWordsTest)
|
|
{
|
|
constexpr SizeType batchSize = 4;
|
|
constexpr SizeType beamWidth = 2;
|
|
// Expected to match {12, 13} to {bi, bw}={{2, 0}}
|
|
// Expected to match {11, 12} to {bi, bw}={{3, 0}}
|
|
// Expected to match {27} to {bi, bw}={{3, 1}}
|
|
this->runStopWordsCriteriaTest({{{2}}, {{}}, {{11}, {12, 13}}, {{27}, {11, 12}}}, batchSize, beamWidth);
|
|
}
|
|
|
|
} // end of namespace
|