mirror of
https://github.com/NVIDIA/TensorRT-LLM.git
synced 2026-01-14 06:27:45 +08:00
469 lines
15 KiB
C++
469 lines
15 KiB
C++
|
|
/*
|
|
* SPDX-FileCopyrightText: Copyright (c) 1993-2024 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 "lowLatencyGemmSwigluPlugin.h"
|
|
#include "low_latency_gemm_swiglu.h"
|
|
#include "tensorrt_llm/common/assert.h"
|
|
#include "tensorrt_llm/common/cudaFp8Utils.h"
|
|
#include "tensorrt_llm/common/logger.h"
|
|
#include <NvInferRuntime.h>
|
|
#include <NvInferRuntimeBase.h>
|
|
#include <NvInferRuntimePlugin.h>
|
|
#include <cstddef>
|
|
#include <cstdint>
|
|
#include <cstdio>
|
|
#include <numeric>
|
|
#include <optional>
|
|
#include <vector>
|
|
|
|
using namespace nvinfer1;
|
|
using namespace tensorrt_llm::common;
|
|
using namespace tensorrt_llm::kernels::internal_cutlass_kernels;
|
|
using tensorrt_llm::plugins::LowLatencyGemmSwigluPluginCreator;
|
|
using tensorrt_llm::plugins::LowLatencyGemmSwigluPlugin;
|
|
using tensorrt_llm::plugins::LowLatencyGemmSwigluPluginProfiler;
|
|
using tensorrt_llm::plugins::read;
|
|
using tensorrt_llm::plugins::write;
|
|
|
|
static char const* LOW_LATENCY_GEMM_SWIGLU_PLUGIN_VERSION{"1"};
|
|
static char const* LOW_LATENCY_GEMM_SWIGLU_PLUGIN_NAME{"LowLatencyGemmSwiglu"};
|
|
|
|
PluginFieldCollection LowLatencyGemmSwigluPluginCreator::mFC{};
|
|
std::vector<nvinfer1::PluginField> LowLatencyGemmSwigluPluginCreator::mPluginAttributes;
|
|
|
|
using FP8Type = __nv_fp8_e4m3;
|
|
|
|
static std::optional<float> getFloatEnv(char const* name)
|
|
{
|
|
char const* const env = std::getenv(name);
|
|
if (env == nullptr)
|
|
{
|
|
return std::nullopt;
|
|
}
|
|
try
|
|
{
|
|
float value = std::stof(env);
|
|
return {value};
|
|
}
|
|
catch (std::invalid_argument const& e)
|
|
{
|
|
return std::nullopt;
|
|
}
|
|
catch (std::out_of_range const& e)
|
|
{
|
|
return std::nullopt;
|
|
}
|
|
};
|
|
|
|
static size_t getBytePerElement(nvinfer1::DataType type)
|
|
{
|
|
size_t bpe;
|
|
if (type == nvinfer1::DataType::kFLOAT)
|
|
{
|
|
bpe = 4;
|
|
}
|
|
else if (type == nvinfer1::DataType::kHALF || type == nvinfer1::DataType::kBF16)
|
|
{
|
|
bpe = 2;
|
|
}
|
|
else if (type == nvinfer1::DataType::kINT8 || type == nvinfer1::DataType::kFP8)
|
|
{
|
|
bpe = 1;
|
|
}
|
|
else
|
|
{
|
|
TLLM_THROW("Not recognized/implemented");
|
|
}
|
|
return bpe;
|
|
}
|
|
|
|
void LowLatencyGemmSwigluPluginProfiler::runTactic(int m, int n, int k,
|
|
LowLatencyGemmSwigluPluginProfiler::Config const& tactic, char* workspace, cudaStream_t const& stream)
|
|
{
|
|
|
|
float default_pdl_overlap_ratio = 0.5;
|
|
float default_prefetch_ratio = -1.0;
|
|
FP8Type* aTmp = reinterpret_cast<FP8Type*>(workspace);
|
|
FP8Type* bTmp
|
|
= reinterpret_cast<FP8Type*>(nextWorkspacePtr(reinterpret_cast<int8_t*>(aTmp), m * k * sizeof(FP8Type)));
|
|
void* dTmp = reinterpret_cast<void*>(nextWorkspacePtr(reinterpret_cast<int8_t*>(bTmp), n * k * sizeof(FP8Type)));
|
|
size_t workspaceSize = mRunner->getWorkspaceSize(m, n, k);
|
|
char* workspaceTmp = reinterpret_cast<char*>(
|
|
nextWorkspacePtr(reinterpret_cast<int8_t*>(dTmp), (n / 2 * m * getBytePerElement(mType))));
|
|
mRunner->gemm(aTmp, bTmp, 1.0f, 0.0f, 1.0f, 1.0f, nullptr, dTmp, m, n, k, default_pdl_overlap_ratio,
|
|
default_prefetch_ratio, tactic, workspaceTmp, workspaceSize, stream);
|
|
}
|
|
|
|
int LowLatencyGemmSwigluPluginProfiler::getMaxProfileM() const
|
|
{
|
|
return 32768;
|
|
}
|
|
|
|
void LowLatencyGemmSwigluPluginProfiler::computeTmpSize(size_t maxM, size_t n, size_t k)
|
|
{
|
|
|
|
std::vector<size_t> workspaces = {maxM * k * sizeof(FP8Type), // A
|
|
n * k * sizeof(FP8Type), // B
|
|
maxM * (n / 2) * getBytePerElement(mType), // D
|
|
mRunner->getWorkspaceSize(maxM, n, k)}; // workspace
|
|
|
|
size_t bytes = calculateTotalWorkspaceSize(workspaces.data(), workspaces.size());
|
|
setTmpWorkspaceSizeInBytes(bytes);
|
|
}
|
|
|
|
std::vector<LowLatencyGemmSwigluPluginProfiler::Config> LowLatencyGemmSwigluPluginProfiler::getTactics(
|
|
int m, int n, int k) const
|
|
{
|
|
return mRunner->getConfigs();
|
|
}
|
|
|
|
LowLatencyGemmSwigluPlugin::LowLatencyGemmSwigluPlugin(nvinfer1::DataType type, float scale_output, float scale_d0,
|
|
float scale_d1, PluginProfilerPtr const& pluginProfiler)
|
|
: mPluginProfiler(pluginProfiler)
|
|
, mScaleOutput(scale_output)
|
|
, mScaleD0(scale_d0)
|
|
, mScaleD1(scale_d1)
|
|
{
|
|
init(type);
|
|
}
|
|
|
|
LowLatencyGemmSwigluPlugin::LowLatencyGemmSwigluPlugin(
|
|
void const* data, size_t length, PluginProfilerPtr const& pluginProfiler)
|
|
: mPluginProfiler(pluginProfiler)
|
|
{
|
|
|
|
char const *d = reinterpret_cast<char const*>(data), *a = d;
|
|
nvinfer1::DataType type;
|
|
read(d, type);
|
|
read(d, mScaleOutput);
|
|
read(d, mScaleD0);
|
|
read(d, mScaleD1);
|
|
read(d, mDims);
|
|
|
|
init(type);
|
|
mPluginProfiler->deserialize(d, mDims, mGemmId);
|
|
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 LowLatencyGemmSwigluPlugin::init(nvinfer1::DataType type)
|
|
{
|
|
|
|
mType = type;
|
|
|
|
if (mType == nvinfer1::DataType::kFP8)
|
|
{
|
|
mLowLatencyGemmSwigluRunner = std::make_shared<CutlassLowLatencyFp8GemmSwigluRunner<__nv_fp8_e4m3>>();
|
|
}
|
|
else
|
|
{
|
|
TLLM_THROW("Unsupported data type");
|
|
}
|
|
mGemmId = GemmIdCore(mDims.n, mDims.k, mType);
|
|
}
|
|
|
|
// IPluginV2DynamicExt Methods
|
|
nvinfer1::IPluginV2DynamicExt* LowLatencyGemmSwigluPlugin::clone() const noexcept
|
|
{
|
|
auto* plugin = new LowLatencyGemmSwigluPlugin(*this);
|
|
return plugin;
|
|
}
|
|
|
|
nvinfer1::DimsExprs LowLatencyGemmSwigluPlugin::getOutputDimensions(
|
|
int outputIndex, nvinfer1::DimsExprs const* inputs, int nbInputs, nvinfer1::IExprBuilder& exprBuilder) noexcept
|
|
{
|
|
try
|
|
{
|
|
TLLM_CHECK(nbInputs == 2);
|
|
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] = exprBuilder.constant(inputs[1].d[1]->getConstantValue() / 2);
|
|
return ret;
|
|
}
|
|
catch (std::exception const& e)
|
|
{
|
|
caughtError(e);
|
|
}
|
|
return DimsExprs{};
|
|
}
|
|
|
|
bool LowLatencyGemmSwigluPlugin::supportsFormatCombination(
|
|
int pos, nvinfer1::PluginTensorDesc const* inOut, int nbInputs, int nbOutputs) noexcept
|
|
{
|
|
switch (pos)
|
|
{
|
|
case 0:
|
|
// activation
|
|
return inOut[pos].type == nvinfer1::DataType::kFP8 && inOut[pos].format == TensorFormat::kLINEAR;
|
|
case 1:
|
|
// weights
|
|
// Weights stored in checkpoint must have fp8 type
|
|
return inOut[pos].type == nvinfer1::DataType::kFP8 && inOut[pos].format == TensorFormat::kLINEAR;
|
|
case 2:
|
|
// out
|
|
return inOut[pos].type == mType && inOut[pos].format == TensorFormat::kLINEAR;
|
|
default:
|
|
// Never should be here
|
|
TLLM_CHECK(false);
|
|
return false;
|
|
}
|
|
}
|
|
|
|
void LowLatencyGemmSwigluPlugin::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[1];
|
|
int const minK = in[0].min.d[in[0].min.nbDims - 1];
|
|
int const minN = in[1].min.d[1];
|
|
|
|
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};
|
|
}
|
|
mGemmId = {maxN, maxK, mType};
|
|
|
|
mWorkspaceMaxSize = mLowLatencyGemmSwigluRunner->getWorkspaceSize(maxM, maxN, maxK);
|
|
}
|
|
|
|
size_t LowLatencyGemmSwigluPlugin::getWorkspaceSize(nvinfer1::PluginTensorDesc const* inputs, int nbInputs,
|
|
nvinfer1::PluginTensorDesc const* outputs, int nbOutputs) const noexcept
|
|
{
|
|
return mWorkspaceMaxSize;
|
|
}
|
|
|
|
int LowLatencyGemmSwigluPlugin::enqueue(nvinfer1::PluginTensorDesc const* inputDesc,
|
|
nvinfer1::PluginTensorDesc const* outputDesc, void const* const* inputs, void* const* outputs, void* workspace,
|
|
cudaStream_t stream) noexcept
|
|
{
|
|
|
|
// input0 activation [M,K] row-major
|
|
// input1 weights [K, N] col-major
|
|
// output0 [M,N / 2] row-major
|
|
|
|
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[1]);
|
|
int const k = TLLM_INT32_CAST(inputDesc[0].dims.d[inputDesc[0].dims.nbDims - 1]);
|
|
int const wsSize = mLowLatencyGemmSwigluRunner->getWorkspaceSize(m, n, k);
|
|
auto const& bestTactic = mPluginProfiler->getBestConfig(m, mGemmId);
|
|
TLLM_CHECK_WITH_INFO(bestTactic, "No valid Low Latency GEMM SWIGLU tactic");
|
|
|
|
auto env_pdl_overlap_ratio = getFloatEnv("TRTLLM_PDL_OVERLAP_RATIO");
|
|
auto env_prefetch_ratio = getFloatEnv("TRTLLM_PREFETCH_RATIO");
|
|
auto valid_ratio = [](std::optional<float>& env_val, float default_val)
|
|
{
|
|
if (env_val.has_value())
|
|
{
|
|
TLLM_CHECK_WITH_INFO(env_val.value() <= 1.0f, "Valid ratio should be less than or equal to 1.0");
|
|
return env_val.value();
|
|
}
|
|
return default_val;
|
|
};
|
|
float pdl_overlap_ratio = valid_ratio(env_pdl_overlap_ratio, /*default_val=*/0.5);
|
|
float prefetch_ratio = valid_ratio(env_prefetch_ratio, /*default_val=*/-1.0);
|
|
mLowLatencyGemmSwigluRunner->gemm(const_cast<FP8Type*>(reinterpret_cast<FP8Type const*>(inputs[0])),
|
|
const_cast<FP8Type*>(reinterpret_cast<FP8Type const*>(inputs[1])), mScaleOutput, 0.0F, mScaleD0, mScaleD1,
|
|
nullptr, outputs[0], m, n, k, pdl_overlap_ratio, prefetch_ratio, *bestTactic,
|
|
reinterpret_cast<char*>(workspace), wsSize, stream);
|
|
|
|
return 0;
|
|
}
|
|
|
|
// IPluginV2Ext Methods
|
|
nvinfer1::DataType LowLatencyGemmSwigluPlugin::getOutputDataType(
|
|
int index, nvinfer1::DataType const* inputTypes, int nbInputs) const noexcept
|
|
{
|
|
TLLM_CHECK(index == 0);
|
|
return mType;
|
|
}
|
|
|
|
// IPluginV2 Methods
|
|
|
|
char const* LowLatencyGemmSwigluPlugin::getPluginType() const noexcept
|
|
{
|
|
return LOW_LATENCY_GEMM_SWIGLU_PLUGIN_NAME;
|
|
}
|
|
|
|
char const* LowLatencyGemmSwigluPlugin::getPluginVersion() const noexcept
|
|
{
|
|
return LOW_LATENCY_GEMM_SWIGLU_PLUGIN_VERSION;
|
|
}
|
|
|
|
int LowLatencyGemmSwigluPlugin::getNbOutputs() const noexcept
|
|
{
|
|
return 1;
|
|
}
|
|
|
|
int LowLatencyGemmSwigluPlugin::initialize() noexcept
|
|
{
|
|
configGemm();
|
|
return 0;
|
|
}
|
|
|
|
void LowLatencyGemmSwigluPlugin::terminate() noexcept {}
|
|
|
|
size_t LowLatencyGemmSwigluPlugin::getSerializationSize() const noexcept
|
|
{
|
|
return sizeof(nvinfer1::DataType) + // dtype
|
|
sizeof(float) * 3 + // scales
|
|
sizeof(mDims) + mPluginProfiler->getSerializationSize(mGemmId);
|
|
}
|
|
|
|
void LowLatencyGemmSwigluPlugin::serialize(void* buffer) const noexcept
|
|
{
|
|
char *d = static_cast<char*>(buffer), *a = d;
|
|
write(d, mType);
|
|
write(d, mScaleOutput);
|
|
write(d, mScaleD0);
|
|
write(d, mScaleD1);
|
|
write(d, mDims);
|
|
mPluginProfiler->serialize(d, mGemmId);
|
|
TLLM_CHECK(d == a + getSerializationSize());
|
|
}
|
|
|
|
void LowLatencyGemmSwigluPlugin::destroy() noexcept
|
|
{
|
|
// This gets called when the network containing plugin is destroyed
|
|
delete this;
|
|
}
|
|
|
|
void LowLatencyGemmSwigluPlugin::configGemm()
|
|
{
|
|
mPluginProfiler->profileTactics(mLowLatencyGemmSwigluRunner, mType, mDims, mGemmId);
|
|
}
|
|
|
|
//////////////////////////////////////////////////////////////////////////
|
|
|
|
LowLatencyGemmSwigluPluginCreator::LowLatencyGemmSwigluPluginCreator()
|
|
{
|
|
|
|
// Fill PluginFieldCollection with PluginField arguments metadata
|
|
mPluginAttributes.clear();
|
|
mPluginAttributes.emplace_back(PluginField("type_id", nullptr, PluginFieldType::kINT32));
|
|
mPluginAttributes.emplace_back(PluginField("scale_output", nullptr, PluginFieldType::kFLOAT32));
|
|
mPluginAttributes.emplace_back(PluginField("scale_d0", nullptr, PluginFieldType::kFLOAT32));
|
|
mPluginAttributes.emplace_back(PluginField("scale_d1", nullptr, PluginFieldType::kFLOAT32));
|
|
mFC.nbFields = mPluginAttributes.size();
|
|
mFC.fields = mPluginAttributes.data();
|
|
}
|
|
|
|
char const* LowLatencyGemmSwigluPluginCreator::getPluginName() const noexcept
|
|
{
|
|
return LOW_LATENCY_GEMM_SWIGLU_PLUGIN_NAME;
|
|
}
|
|
|
|
char const* LowLatencyGemmSwigluPluginCreator::getPluginVersion() const noexcept
|
|
{
|
|
return LOW_LATENCY_GEMM_SWIGLU_PLUGIN_VERSION;
|
|
}
|
|
|
|
PluginFieldCollection const* LowLatencyGemmSwigluPluginCreator::getFieldNames() noexcept
|
|
{
|
|
return &mFC;
|
|
}
|
|
|
|
IPluginV2* LowLatencyGemmSwigluPluginCreator::createPlugin(char const* name, PluginFieldCollection const* fc) noexcept
|
|
{
|
|
PluginField const* fields = fc->fields;
|
|
TLLM_CHECK(fc->nbFields == 4);
|
|
nvinfer1::DataType type{};
|
|
float scale_output{};
|
|
float scale_d0{};
|
|
float scale_d1{};
|
|
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);
|
|
type = static_cast<nvinfer1::DataType>(*(static_cast<nvinfer1::DataType const*>(fields[i].data)));
|
|
}
|
|
else if (!strcmp(attrName, "scale_output"))
|
|
{
|
|
TLLM_CHECK(fields[i].type == PluginFieldType::kFLOAT32);
|
|
scale_output = static_cast<float>(*(static_cast<float const*>(fields[i].data)));
|
|
}
|
|
else if (!strcmp(attrName, "scale_d0"))
|
|
{
|
|
TLLM_CHECK(fields[i].type == PluginFieldType::kFLOAT32);
|
|
scale_d0 = static_cast<float>(*(static_cast<float const*>(fields[i].data)));
|
|
}
|
|
else if (!strcmp(attrName, "scale_d1"))
|
|
{
|
|
TLLM_CHECK(fields[i].type == PluginFieldType::kFLOAT32);
|
|
scale_d1 = static_cast<float>(*(static_cast<float const*>(fields[i].data)));
|
|
}
|
|
}
|
|
|
|
try
|
|
{
|
|
|
|
//
|
|
// LowLatencyGemmSwigluPluginCreator is unique and shared for an engine generation
|
|
// Create plugin profiler with shared tactics map
|
|
auto pluginProfiler = gemmPluginProfileManager.createGemmPluginProfiler(/*inference=*/false);
|
|
auto* obj = new LowLatencyGemmSwigluPlugin(type, scale_output, scale_d0, scale_d1, pluginProfiler);
|
|
obj->setPluginNamespace(mNamespace.c_str());
|
|
return obj;
|
|
}
|
|
|
|
catch (std::exception const& e)
|
|
{
|
|
caughtError(e);
|
|
}
|
|
return nullptr;
|
|
}
|
|
|
|
IPluginV2* LowLatencyGemmSwigluPluginCreator::deserializePlugin(
|
|
char const* name, void const* serialData, size_t serialLength) noexcept
|
|
{
|
|
try
|
|
{
|
|
auto pluginProfiler = gemmPluginProfileManager.createGemmPluginProfiler(/*inference=*/true);
|
|
auto* obj = new LowLatencyGemmSwigluPlugin(serialData, serialLength, pluginProfiler);
|
|
obj->setPluginNamespace(mNamespace.c_str());
|
|
return obj;
|
|
}
|
|
catch (std::exception const& e)
|
|
{
|
|
caughtError(e);
|
|
}
|
|
return nullptr;
|
|
}
|