TensorRT-LLMs/cpp/tensorrt_llm/plugins/qserveGemmPlugin/qserveGemmPlugin.cpp
Kaiyu Xie 2ea17cdad2
Update TensorRT-LLM (#2792)
* Update TensorRT-LLM

---------

Co-authored-by: jlee <jungmoolee@clika.io>
2025-02-18 21:27:39 +08:00

417 lines
14 KiB
C++

/*
* SPDX-FileCopyrightText: Copyright (c) 1993-2022 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 "qserveGemmPlugin.h"
#include "tensorrt_llm/kernels/qserveGemm.h"
#include <cassert>
#include <numeric>
using namespace nvinfer1;
using namespace tensorrt_llm::common;
using tensorrt_llm::plugins::QServeGemmPluginCreator;
using tensorrt_llm::plugins::QServeGemmPlugin;
using tensorrt_llm::plugins::read;
using tensorrt_llm::plugins::write;
using namespace tensorrt_llm::kernels::qserve;
static char const* QSERVE_GEMM_PLUGIN_VERSION{"1"};
static char const* QSERVE_GEMM_PLUGIN_NAME{"QServeGemm"};
PluginFieldCollection QServeGemmPluginCreator::mFC{};
std::vector<nvinfer1::PluginField> QServeGemmPluginCreator::mPluginAttributes;
namespace tensorrt_llm::plugins
{
QServeGemmPlugin::QServeGemmPlugin(
// QuantMode quantMode,
nvinfer1::DataType dtype, int groupSize)
{
init(dtype, groupSize);
}
QServeGemmPlugin::QServeGemmPlugin(void const* data, size_t length)
{
char const *d = reinterpret_cast<char const*>(data), *a = d;
nvinfer1::DataType type;
unsigned int quantMode;
int groupSize;
read(d, quantMode);
read(d, type);
read(d, groupSize);
read(d, mDims);
// mQuantMode = QuantMode(quantMode);
init(type, groupSize);
TLLM_CHECK_WITH_INFO(d == a + length,
"Expected length (%d) != real length (%d). This is often "
"caused by using different TensorRT-LLM version to build "
"engine and run engine.",
(int) length, (int) (d - a));
}
void QServeGemmPlugin::init(nvinfer1::DataType dtype, int groupSize)
{
if (groupSize <= 0)
groupSize = -1; // Per-channel
mGroupSize = groupSize;
mType = dtype;
mRunner = std::make_shared<QServeGemmRunner>();
}
// IPluginV2DynamicExt Methods
nvinfer1::IPluginV2DynamicExt* QServeGemmPlugin::clone() const noexcept
{
auto* plugin = new QServeGemmPlugin(*this);
return plugin;
}
nvinfer1::DimsExprs QServeGemmPlugin::getOutputDimensions(
int outputIndex, nvinfer1::DimsExprs const* inputs, int nbInputs, nvinfer1::IExprBuilder& exprBuilder) noexcept
{
try
{
TLLM_CHECK(nbInputs == 6);
TLLM_CHECK(outputIndex == 0);
int const nbDimsA = inputs[0].nbDims;
TLLM_CHECK(nbDimsA >= 2);
DimsExprs ret;
ret.nbDims = nbDimsA;
for (int ii = 0; ii < nbDimsA - 1; ++ii)
{
ret.d[ii] = inputs[0].d[ii];
}
ret.d[nbDimsA - 1] = inputs[1].d[0];
return ret;
}
catch (std::exception const& e)
{
caughtError(e);
}
return DimsExprs{};
}
bool QServeGemmPlugin::supportsFormatCombination(
int pos, nvinfer1::PluginTensorDesc const* inOut, int nbInputs, int nbOutputs) noexcept
{
if (mGroupSize != -1)
{ // Per-group
switch (pos)
{
case 0:
// activation
return inOut[pos].type == nvinfer1::DataType::kINT8 && inOut[pos].format == TensorFormat::kLINEAR;
case 1:
// uint4 weights packed in int8
return inOut[pos].type == nvinfer1::DataType::kINT8 && inOut[pos].format == TensorFormat::kLINEAR;
case 2:
// int8 weight s2_zeros
return inOut[pos].type == nvinfer1::DataType::kINT8 && inOut[pos].format == TensorFormat::kLINEAR;
case 3:
// int8 weight s2_scales
return inOut[pos].type == nvinfer1::DataType::kINT8 && inOut[pos].format == TensorFormat::kLINEAR;
case 4:
// fp16 weight s1_scales
return inOut[pos].type == nvinfer1::DataType::kHALF && inOut[pos].format == TensorFormat::kLINEAR;
case 5:
// fp16 activation scales
return inOut[pos].type == nvinfer1::DataType::kHALF && inOut[pos].format == TensorFormat::kLINEAR;
case 6:
// fp16 output activation
return inOut[pos].type == nvinfer1::DataType::kHALF && inOut[pos].format == TensorFormat::kLINEAR;
default: return false;
}
}
else
{ // Per-channel
switch (pos)
{
case 0:
// activation
return inOut[pos].type == nvinfer1::DataType::kINT8 && inOut[pos].format == TensorFormat::kLINEAR;
case 1:
// uint4 weights packed in int8
return inOut[pos].type == nvinfer1::DataType::kINT8 && inOut[pos].format == TensorFormat::kLINEAR;
case 2:
// fp16 s1_scales
return inOut[pos].type == nvinfer1::DataType::kHALF && inOut[pos].format == TensorFormat::kLINEAR;
case 3:
// fp16 s1_szeros
return inOut[pos].type == nvinfer1::DataType::kHALF && inOut[pos].format == TensorFormat::kLINEAR;
case 4:
// fp16 act_sums
return inOut[pos].type == nvinfer1::DataType::kHALF && inOut[pos].format == TensorFormat::kLINEAR;
case 5:
// fp16 act_scales
return inOut[pos].type == nvinfer1::DataType::kHALF && inOut[pos].format == TensorFormat::kLINEAR;
case 6:
// fp16 output activation
return inOut[pos].type == nvinfer1::DataType::kHALF && inOut[pos].format == TensorFormat::kLINEAR;
default: return false;
}
}
}
void QServeGemmPlugin::configurePlugin(nvinfer1::DynamicPluginTensorDesc const* in, int nbInputs,
nvinfer1::DynamicPluginTensorDesc const* out, int nbOutputs) noexcept
{
auto const minM = std::accumulate(in[0].min.d, in[0].min.d + in[0].min.nbDims - 1, 1, std::multiplies<int>());
auto const maxM = std::accumulate(in[0].max.d, in[0].max.d + in[0].max.nbDims - 1, 1, std::multiplies<int>());
int const maxK = in[0].max.d[in[0].max.nbDims - 1];
int const maxN = in[1].max.d[0];
int const minK = in[0].min.d[in[0].min.nbDims - 1];
int const minN = in[1].min.d[0];
TLLM_CHECK_WITH_INFO(minN == maxN, "Variable out channels is not allowed");
TLLM_CHECK_WITH_INFO(minK == maxK, "Variable in channels is not allowed");
if (!mDims.isInitialized())
{
mDims = {minM, maxM, maxN, maxK};
}
m_workspaceMaxSize = mRunner->getWorkspaceSize(maxM, maxN, maxK);
}
size_t QServeGemmPlugin::getWorkspaceSize(nvinfer1::PluginTensorDesc const* inputs, int nbInputs,
nvinfer1::PluginTensorDesc const* outputs, int nbOutputs) const noexcept
{
return m_workspaceMaxSize;
}
int QServeGemmPlugin::enqueue(nvinfer1::PluginTensorDesc const* inputDesc, nvinfer1::PluginTensorDesc const* outputDesc,
void const* const* inputs, void* const* outputs, void* workspace, cudaStream_t stream) noexcept
{
// inputs
// Per group:
// activation [M, K] int8_t Quantized sint8 activations
// weights [N, K/2] int8_t Quantized uint4 weights (packed as int8_t)
// s2_zeros [K/group_size, N] int8_t Level-2 sint8 scaled zeros of weights
// s2_scales [K/group_size, N] int8_t Level-2 sint8 scales of weights
// s1_scales [N] half Level-1 fp16 scales of weights
// act_scales [M] half Scales of activations
// Per channel:
// activation [M, K] int8_t Quantized sint8 activations
// weights [N, K/2] int8_t Quantized uint4 weights (packed as int8_t)
// s1_scales [N] half Level-1 scales of weights
// s1_szeros [N] half Level-1 scaled zeros of weights
// act_sums [M] half Per-token sums of activations
// act_scales [M] half Scales of activations
// outputs
// mat [M(*), N] half
int64_t m64 = 1;
for (int ii = 0; ii < inputDesc[0].dims.nbDims - 1; ++ii)
{
m64 *= inputDesc[0].dims.d[ii];
}
int const m = TLLM_INT32_CAST(m64);
int const n = TLLM_INT32_CAST(inputDesc[1].dims.d[0]);
int const k = TLLM_INT32_CAST(inputDesc[0].dims.d[inputDesc[0].dims.nbDims - 1]);
// TODO: Implement optimized kernels if (m <= 4)
if (mGroupSize != -1)
{
ParamsPerGroup params = {reinterpret_cast<int8_t const*>(inputs[0]), // A
reinterpret_cast<int8_t const*>(inputs[1]), // B
reinterpret_cast<int8_t const*>(inputs[2]), // s2_zeros
reinterpret_cast<int8_t const*>(inputs[3]), // s2_scales
reinterpret_cast<half const*>(inputs[4]), // s1_scales
reinterpret_cast<half const*>(inputs[5]), // act_scales
reinterpret_cast<half*>(outputs[0]), // C
m, n, k};
mRunner->gemmPerGroup(params, stream);
}
else
{
ParamsPerChannel params = {reinterpret_cast<int8_t const*>(inputs[0]), // A
reinterpret_cast<int8_t const*>(inputs[1]), // B
reinterpret_cast<half const*>(inputs[2]), // s1_scales
reinterpret_cast<half const*>(inputs[3]), // s1_szeros
reinterpret_cast<half const*>(inputs[4]), // act_sums
reinterpret_cast<half const*>(inputs[5]), // act_scales
reinterpret_cast<half*>(outputs[0]), // C
m, n, k};
mRunner->gemmPerChannel(params, stream);
}
return 0;
}
// IPluginV2Ext Methods
nvinfer1::DataType QServeGemmPlugin::getOutputDataType(
int index, nvinfer1::DataType const* inputTypes, int nbInputs) const noexcept
{
TLLM_CHECK(index == 0);
return mType;
}
// IPluginV2 Methods
char const* QServeGemmPlugin::getPluginType() const noexcept
{
return QSERVE_GEMM_PLUGIN_NAME;
}
char const* QServeGemmPlugin::getPluginVersion() const noexcept
{
return QSERVE_GEMM_PLUGIN_VERSION;
}
int QServeGemmPlugin::getNbOutputs() const noexcept
{
return 1;
}
int QServeGemmPlugin::initialize() noexcept
{
configGemm();
return 0;
}
void QServeGemmPlugin::terminate() noexcept {}
size_t QServeGemmPlugin::getSerializationSize() const noexcept
{
return sizeof(mQuantMode) + // QuantMode
sizeof(mType) + // dtype
sizeof(mGroupSize) + // GroupSize
sizeof(mDims); // Dimensions
}
void QServeGemmPlugin::serialize(void* buffer) const noexcept
{
char *d = static_cast<char*>(buffer), *a = d;
write(d, mQuantMode.value());
write(d, mType);
write(d, mGroupSize);
write(d, mDims);
TLLM_CHECK(d == a + getSerializationSize());
}
void QServeGemmPlugin::destroy() noexcept
{
// This gets called when the network containing plugin is destroyed
delete this;
}
void QServeGemmPlugin::configGemm() {}
///////////////
QServeGemmPluginCreator::QServeGemmPluginCreator()
{
// Fill PluginFieldCollection with PluginField arguments metadata
mPluginAttributes.clear();
mPluginAttributes.push_back(PluginField("type_id", nullptr, PluginFieldType::kINT32));
mPluginAttributes.push_back(PluginField("group_size", nullptr, PluginFieldType::kINT32));
mFC.nbFields = mPluginAttributes.size();
mFC.fields = mPluginAttributes.data();
}
char const* QServeGemmPluginCreator::getPluginName() const noexcept
{
return QSERVE_GEMM_PLUGIN_NAME;
}
char const* QServeGemmPluginCreator::getPluginVersion() const noexcept
{
return QSERVE_GEMM_PLUGIN_VERSION;
}
PluginFieldCollection const* QServeGemmPluginCreator::getFieldNames() noexcept
{
return &mFC;
}
IPluginV2* QServeGemmPluginCreator::createPlugin(char const* name, PluginFieldCollection const* fc) noexcept
{
// We do not use any fields for now.
PluginField const* fields = fc->fields;
// bool perTokenScaling, perChannelScaling;
DataType dtype{};
int group_size = -1;
// Read configurations from each fields
for (int i = 0; i < fc->nbFields; ++i)
{
char const* attrName = fields[i].name;
if (!strcmp(attrName, "type_id"))
{
TLLM_CHECK(fields[i].type == PluginFieldType::kINT32);
dtype = static_cast<nvinfer1::DataType>(*(static_cast<nvinfer1::DataType const*>(fields[i].data)));
// Only supports fp16 for now.
assert(dtype == nvinfer1::DataType::kHALF);
}
else if (!strcmp(attrName, "group_size"))
{
TLLM_CHECK(fields[i].type == PluginFieldType::kINT32);
group_size = *static_cast<int const*>(fields[i].data);
// Currently only support per-channel or g128.
assert(group_size == -1 || group_size == 128);
}
}
try
{
// QServeGemmPluginCreator is unique and shared for an engine generation
// Create plugin profiler with shared tactics map
// auto pluginProfiler = gemmPluginProfileManager.createGemmPluginProfiler(/* inference */ false);
// QuantMode quantMode = QuantMode::fromQuantAlgo("W4A8_QSERVE");
auto* obj = new QServeGemmPlugin(dtype, group_size);
obj->setPluginNamespace(mNamespace.c_str());
return obj;
}
catch (std::exception const& e)
{
caughtError(e);
}
return nullptr;
}
IPluginV2* QServeGemmPluginCreator::deserializePlugin(
char const* name, void const* serialData, size_t serialLength) noexcept
{
// This object will be deleted when the network is destroyed, which will
// call QServeGemmPlugin::destroy()
try
{
// Create plugin profiler with private tactics map which is read from the serialized engine
// auto pluginProfiler = gemmPluginProfileManager.createGemmPluginProfiler(/* inference */ true);
auto* obj = new QServeGemmPlugin(serialData, serialLength);
obj->setPluginNamespace(mNamespace.c_str());
return obj;
}
catch (std::exception const& e)
{
caughtError(e);
}
return nullptr;
}
} // namespace tensorrt_llm::plugins