TensorRT-LLMs/cpp/tests/unit_tests/kernels/allReduce/allReduceFusionTest.cu
hlu1 8207d5fd39
[None] [feat] Add model gpt-oss (#6645)
Signed-off-by: Hao Lu <14827759+hlu1@users.noreply.github.com>
2025-08-07 03:04:18 -04:00

736 lines
30 KiB
Plaintext

/*
* Copyright (c) 2022-2025, 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.
*/
#include <cuda_runtime.h>
#include <gtest/gtest.h>
#include <nccl.h>
#include <cstdarg>
#include <cstdint>
#include <functional>
#include <iostream>
#include <random>
#include <vector>
#include "tensorrt_llm/kernels/communicationKernels/allReduceFusionKernels.h"
#include "tensorrt_llm/kernels/communicationKernels/allReduceWorkspace.h"
#include "tensorrt_llm/kernels/quantization.h"
#include "tensorrt_llm/kernels/rmsnormKernels.h"
#include "tensorrt_llm/runtime/cudaStream.h"
#include "tensorrt_llm/runtime/utils/mpiUtils.h"
#include "tensorrt_llm/runtime/utils/multiDeviceUtils.h"
namespace mpi = tensorrt_llm::mpi;
namespace tr = tensorrt_llm::runtime;
using namespace tensorrt_llm::kernels;
template <typename DType>
__global__ void residual_add_kernel(DType* data, DType* residual, int size)
{
int idx = blockIdx.x * blockDim.x + threadIdx.x;
if (idx >= size)
return;
data[idx] = data[idx] + residual[idx];
}
template <typename DType>
void residual_add(DType* data, DType* residual, int size, cudaStream_t stream)
{
residual_add_kernel<<<(size + 127) / 128, 128, 0, stream>>>(data, residual, size);
}
template <typename DType>
__global__ void quantize_to_fp8_kernel(DType* data, __nv_fp8_e4m3* data_fp8, int size, float* scale_factor)
{
int idx = blockIdx.x * blockDim.x + threadIdx.x;
if (idx >= size)
return;
data_fp8[idx] = static_cast<__nv_fp8_e4m3>(static_cast<float>(data[idx]) * (1.f / *scale_factor));
}
template <typename DType>
void quantize_to_fp8(DType* data, __nv_fp8_e4m3* data_fp8, int size, float* scale_factor, cudaStream_t stream)
{
quantize_to_fp8_kernel<<<(size + 127) / 128, 128, 0, stream>>>(data, data_fp8, size, scale_factor);
}
template <typename T>
bool compare(int rank, void* p_real, void* p_ref, int size, std::string const& cmp_info = "", float atol = 1e-3)
{
auto ptr_real = reinterpret_cast<T*>(p_real);
auto ptr_ref = reinterpret_cast<T*>(p_ref);
float max_diff = 0.f, tot_diff = 0.f;
int error_cnt = 0;
float max_error_value_real = 0.f, max_error_value_ref = 0.f;
static char* ar_debug = std::getenv("AR_DEBUG");
if (ar_debug && rank == 0)
{
printf("TensorReal: [");
for (int n = 0; n < 20; ++n)
{
float v = static_cast<float>(ptr_real[n]);
printf("%f, ", v);
}
printf("...]\n");
printf("TensorRef: [");
for (int n = 0; n < 20; ++n)
{
float v = static_cast<float>(ptr_ref[n]);
printf("%f, ", v);
}
printf("...]\n");
}
int print_cnt = 0;
for (int n = 0; n < size; ++n)
{
float v_real = static_cast<float>(ptr_real[n]);
float v_ref = static_cast<float>(ptr_ref[n]);
float diff = std::abs(v_real - v_ref);
if (diff > max_diff)
{
max_diff = diff;
max_error_value_real = v_real;
max_error_value_ref = v_ref;
}
bool is_error = diff > atol;
if (diff > atol)
{
tot_diff += diff;
++error_cnt;
}
if (ar_debug && is_error && rank == 0 && print_cnt < 20)
{
++print_cnt;
if (rank == 0)
printf("idx %d, v_real %f, v_ref %f\n", n, v_real, v_ref);
}
}
bool pass = error_cnt == 0;
if (!pass && rank == 0)
{
printf(
"[%s] rank %d, atol %8.4f, max absolute diff %8.4f(%8.4f vs %8.4f), avg absolute diff %8.4f, absolute "
"error count %d/%d\n",
cmp_info.c_str(), rank, atol, max_diff, max_error_value_real, max_error_value_ref,
tot_diff / std::max(error_cnt, 1), error_cnt, size);
}
return pass;
}
template <typename T1, typename T2>
void random_fill(T1* data, int size, T2 minv, T2 maxv)
{
static int rseed = 20250227;
std::mt19937 gen(rseed++);
std::uniform_real_distribution<float> dis(static_cast<float>(minv), static_cast<float>(maxv));
for (int i = 0; i < size; ++i)
{
data[i] = static_cast<T1>(dis(gen));
}
}
int get_random_int(int min_v, int max_v)
{
static int rseed = 20250227;
std::mt19937 gen(rseed++);
std::uniform_int_distribution<> dis(min_v, max_v);
return dis(gen);
}
struct CudaBuffer
{
void* m_d_data;
void* m_h_data;
int m_size;
CudaBuffer(int size_in_bytes = 0)
: m_size(size_in_bytes)
, m_d_data(nullptr)
, m_h_data(nullptr)
{
allocate(size_in_bytes);
}
void allocate(int size_in_bytes)
{
if (size_in_bytes == 0)
return;
TLLM_CHECK(m_d_data == nullptr && m_h_data == nullptr);
m_size = size_in_bytes;
TLLM_CUDA_CHECK(cudaMalloc(&m_d_data, m_size));
clear();
m_h_data = malloc(m_size);
}
template <typename T = void>
T* device_data()
{
TLLM_CHECK(m_d_data != nullptr);
return reinterpret_cast<T*>(m_d_data);
}
template <typename T = void>
T* host_data()
{
TLLM_CHECK(m_h_data != nullptr);
d2h();
return reinterpret_cast<T*>(m_h_data);
}
template <typename DType, typename VType>
void random(VType minv, VType maxv)
{
random_fill(reinterpret_cast<DType*>(m_h_data), m_size / sizeof(DType), minv, maxv);
h2d();
}
void clear()
{
TLLM_CUDA_CHECK(cudaMemset(m_d_data, 0, m_size));
}
void h2d()
{
TLLM_CUDA_CHECK(cudaMemcpy(m_d_data, m_h_data, m_size, cudaMemcpyHostToDevice));
}
void d2h()
{
TLLM_CUDA_CHECK(cudaMemcpy(m_h_data, m_d_data, m_size, cudaMemcpyDeviceToHost));
}
~CudaBuffer()
{
if (m_d_data)
{
TLLM_CUDA_CHECK(cudaFree(m_d_data));
}
if (m_h_data)
{
free(m_h_data);
}
}
};
template <typename DType>
struct DTypeTraits;
template <>
struct DTypeTraits<half>
{
static constexpr ncclDataType_t kNCCLDataType = ncclFloat16;
static constexpr nvinfer1::DataType kTRTDataType = nvinfer1::DataType::kHALF;
};
template <>
struct DTypeTraits<__nv_bfloat16>
{
static constexpr ncclDataType_t kNCCLDataType = ncclBfloat16;
static constexpr nvinfer1::DataType kTRTDataType = nvinfer1::DataType::kBF16;
};
template <>
struct DTypeTraits<float>
{
static constexpr ncclDataType_t kNCCLDataType = ncclFloat32;
static constexpr nvinfer1::DataType kTRTDataType = nvinfer1::DataType::kFLOAT;
};
template <typename DType, ar_fusion::AllReduceFusionPattern Pattern>
class TestRunner
{
static constexpr ncclDataType_t kNCCLDataType = DTypeTraits<DType>::kNCCLDataType;
static constexpr nvinfer1::DataType kTRTDataType = DTypeTraits<DType>::kTRTDataType;
static constexpr bool kFP4QuantOutSupport = !std::is_same_v<DType, float>;
static_assert(kFP4QuantOutSupport || Pattern != ar_fusion::AllReduceFusionPattern::kARResidualRMSNormFP4Quant,
"kARResidualRMSNormFP4Quant is not supported for float dtype");
public:
TestRunner(int max_token_num, int hidden_dim)
: m_mpi_comm(mpi::MpiComm::world())
{
m_message_size = max_token_num * hidden_dim;
m_world_size = m_mpi_comm.getSize();
m_rank = m_mpi_comm.getRank();
TLLM_CUDA_CHECK(cudaSetDevice(m_rank));
ncclUniqueId id;
if (m_rank == 0)
{
TLLM_NCCL_CHECK(ncclGetUniqueId(&id));
}
m_mpi_comm.bcast(&id, sizeof(id), mpi::MpiType::kBYTE, 0);
TLLM_NCCL_CHECK(ncclCommInitRank(&m_nccl_comm, m_world_size, id, m_rank));
m_allreduce_in.allocate(m_message_size * sizeof(DType));
m_residual_in.allocate(m_message_size * sizeof(DType));
m_allreduce_out.allocate(m_message_size * sizeof(DType));
m_residual_out.allocate(m_message_size * sizeof(DType));
m_norm_out.allocate(m_message_size * sizeof(DType));
m_quant_out.allocate(m_message_size * sizeof(DType));
// SF layout was packed to [numMTiles, numKTiles, 32 (mTile), 4 (mTile), 4(kTile)]
size_t scale_out_size = ((max_token_num + 127) / 128 * 128) * ((hidden_dim + 63) / 64 * 4);
m_scale_out.allocate(scale_out_size);
m_rms_gamma.allocate(hidden_dim * sizeof(DType));
m_scale_factor.allocate(sizeof(float));
m_stream = std::make_shared<tr::CudaStream>();
m_workspace = std::make_shared<ar_fusion::Workspace>(m_rank, m_world_size, max_token_num, hidden_dim, m_stream);
m_params.nranks = m_world_size;
m_params.rank = m_rank;
m_params.dtype = kTRTDataType;
m_params.workspace = m_workspace->get_workspace();
m_params.allreduce_in = m_allreduce_in.device_data();
m_params.residual_in = m_residual_in.device_data();
m_params.allreduce_out = m_allreduce_out.device_data();
m_params.residual_out = m_residual_out.device_data();
m_params.norm_out = m_norm_out.device_data();
m_params.quant_out = m_quant_out.device_data();
m_params.scale_out = m_scale_out.device_data();
m_params.rms_gamma = m_rms_gamma.device_data();
m_params.scale_factor = m_scale_factor.device_data<float>();
m_params.rms_eps = 1e-3;
m_params.stream = m_stream->get();
m_params.pattern = Pattern;
}
void reset_io()
{
m_allreduce_in.random<DType>(-100.f, 100.f);
m_residual_in.random<DType>(-100.f, 100.f);
m_rms_gamma.random<DType>(-1.f, 1.f);
m_scale_factor.random<float>(1.f, 1.f);
// Because scale_out internally performs layout interleaving, not all elements will be covered, so it should be
// reset before calling the kernel to ensure correct comparison results
if (kFP4QuantOutSupport)
{
m_scale_out.clear();
}
}
template <typename Func>
float benchmark(Func func, int warmup, int iter, int token_num, int hidden_dim)
{
m_params.size = token_num * hidden_dim;
m_params.hidden_dim = hidden_dim;
cudaEvent_t begin, end;
cudaEventCreate(&begin);
cudaEventCreate(&end);
m_mpi_comm.barrier();
for (int i = 0; i < warmup; ++i)
{
(this->*func)(token_num, hidden_dim);
}
cudaEventRecord(begin, m_stream->get());
for (int i = 0; i < iter; ++i)
{
(this->*func)(token_num, hidden_dim);
}
cudaEventRecord(end, m_stream->get());
cudaEventSynchronize(end);
float time;
cudaEventElapsedTime(&time, begin, end);
time /= iter;
m_mpi_comm.barrier();
cudaEventDestroy(begin);
cudaEventDestroy(end);
return time * 1000;
}
template <typename Func>
void run_once(Func func, int token_num, int hidden_dim)
{
benchmark(func, 0, 1, token_num, hidden_dim);
}
int get_sm_count()
{
static int sm_count = 0;
if (sm_count == 0)
{
int device_id;
TLLM_CUDA_CHECK(cudaGetDevice(&device_id));
cudaDeviceProp device_prop;
cudaGetDeviceProperties(&device_prop, device_id);
sm_count = device_prop.multiProcessorCount;
}
return sm_count;
}
void verify(int token_num, int hidden_dim)
{
int message_size = token_num * hidden_dim;
CudaBuffer ref_output(message_size * sizeof(DType));
// We directly compare the results of AR+AddResidual here, as the accumulation order in NCCL's AR might be
// inconsistent across different kernels. Therefore, we set atol to 1 (setting it to 0 locally also passes the
// test).
TLLM_NCCL_CHECK(ncclAllReduce(m_allreduce_in.device_data(), ref_output.device_data(), message_size,
kNCCLDataType, ncclSum, m_nccl_comm, 0));
if constexpr (ar_fusion::HasAllReduceOut<Pattern>)
{
TLLM_CHECK(compare<DType>(
m_rank, m_allreduce_out.host_data(), ref_output.host_data(), message_size, "allreduce out", 1));
}
if constexpr (ar_fusion::HasResidual<Pattern>)
{
residual_add(ref_output.device_data<DType>(), m_residual_in.device_data<DType>(), message_size, 0);
if constexpr (ar_fusion::HasResidualOut<Pattern>)
{
TLLM_CHECK(compare<DType>(
m_rank, m_residual_out.host_data(), ref_output.host_data(), message_size, "residual out", 1));
}
}
if constexpr (ar_fusion::HasRMSNorm<Pattern>)
{
// This excludes the accumulation order errors introduced by AR and only compares the accuracy of the
// RMSNorm. The atol is set to 1e-2 to exclude errors caused by accumulation order changes due to
// differences in cluster/block size.
invokeGeneralRmsNorm<DType, int8_t>(ref_output.device_data<DType>(), m_residual_out.device_data<DType>(),
m_rms_gamma.device_data<DType>(), nullptr, m_params.rms_eps, token_num, hidden_dim,
tensorrt_llm::common::QuantMode(), 0);
if constexpr (ar_fusion::HasNormOut<Pattern>)
{
TLLM_CHECK(compare<DType>(
m_rank, m_norm_out.host_data(), ref_output.host_data(), message_size, "norm out", 1e-2));
}
}
if constexpr (ar_fusion::GetQuantType<Pattern> == ar_fusion::QuantType::kFP4)
{
// We need norm out to verify the accuracy of quantization.
static_assert(Pattern == ar_fusion::AllReduceFusionPattern::kARResidualRMSNormOutFP4Quant);
// SF layout was packed to [numMTiles, numKTiles, 32 (mTile), 4 (mTile), 4(kTile)]
size_t scale_out_size = ((token_num + 127) / 128 * 128) * ((hidden_dim + 63) / 64 * 4);
CudaBuffer ref_scale(scale_out_size);
// Here, we also only compare the accuracy of quantization. Since there are no differences in
// computation order, atol is set to 0.
invokeFP4Quantization(1, token_num, hidden_dim, m_norm_out.device_data<DType>(),
m_scale_factor.device_data<float>(), ref_output.device_data<int64_t>(),
ref_scale.device_data<int32_t>(), false, tensorrt_llm::QuantizationSFLayout::SWIZZLED, 128, 0);
TLLM_CHECK(compare<int8_t>(
m_rank, m_quant_out.host_data(), ref_output.host_data(), message_size / 2, "fp4 quant out", 0));
TLLM_CHECK(compare<int8_t>(
m_rank, m_scale_out.host_data(), ref_scale.host_data(), scale_out_size, "fp4 scale out", 0));
}
else if constexpr (ar_fusion::GetQuantType<Pattern> == ar_fusion::QuantType::kFP8)
{
// We need norm out to verify the accuracy of quantization.
static_assert(Pattern == ar_fusion::AllReduceFusionPattern::kARResidualRMSNormOutFP8Quant);
CudaBuffer ref_fp8_output(message_size * sizeof(__nv_fp8_e4m3));
quantize_to_fp8(m_norm_out.device_data<DType>(), ref_fp8_output.device_data<__nv_fp8_e4m3>(), message_size,
m_scale_factor.device_data<float>(), m_stream->get());
TLLM_CHECK(compare<__nv_fp8_e4m3>(
m_rank, m_quant_out.host_data(), ref_fp8_output.host_data(), message_size, "fp8 quant out", 0));
}
}
void run_nccl_allreduce(int token_num, int hidden_dim)
{
TLLM_NCCL_CHECK(ncclAllReduce(m_allreduce_in.device_data(), m_residual_out.device_data(),
token_num * hidden_dim, kNCCLDataType, ncclSum, m_nccl_comm, m_stream->get()));
}
void run_residual_add(int token_num, int hidden_dim)
{
residual_add(m_residual_out.device_data<DType>(), m_residual_in.device_data<DType>(), token_num * hidden_dim,
m_stream->get());
}
void run_rms_norm(int token_num, int hidden_dim)
{
invokeGeneralRmsNorm<DType, int8_t>(m_norm_out.device_data<DType>(), m_residual_out.device_data<DType>(),
m_rms_gamma.device_data<DType>(), nullptr, m_params.rms_eps, token_num, hidden_dim,
tensorrt_llm::common::QuantMode(), m_stream->get());
}
void run_fp4_quant(int token_num, int hidden_dim)
{
invokeFP4Quantization(1, token_num, hidden_dim, m_norm_out.device_data<DType>(),
m_scale_factor.device_data<float>(), m_quant_out.device_data<int64_t>(), m_scale_out.device_data<int32_t>(),
false, tensorrt_llm::QuantizationSFLayout::SWIZZLED, 128, m_stream->get());
}
void run_kernel(int token_num, int hidden_dim)
{
ar_fusion::allreduce_fusion_op(m_params);
}
~TestRunner()
{
TLLM_NCCL_CHECK(ncclCommDestroy(m_nccl_comm));
}
private:
int m_rank;
int m_world_size;
int m_message_size;
mpi::MpiComm const& m_mpi_comm;
ncclComm_t m_nccl_comm;
CudaBuffer m_allreduce_in;
CudaBuffer m_residual_in;
CudaBuffer m_allreduce_out;
CudaBuffer m_residual_out;
CudaBuffer m_norm_out;
CudaBuffer m_quant_out;
CudaBuffer m_scale_out;
CudaBuffer m_rms_gamma;
CudaBuffer m_scale_factor;
std::shared_ptr<ar_fusion::Workspace> m_workspace;
ar_fusion::AllReduceFusionParams m_params;
std::shared_ptr<tr::CudaStream> m_stream;
};
TEST(Kernel_AllReduceFusion, AllReduceAccuracyRandomTokenNum)
{
using Runner = TestRunner<half, ar_fusion::AllReduceFusionPattern::kAllReduce>;
auto& comm = mpi::MpiComm::world();
auto world_size = comm.getSize();
auto rank = comm.getRank();
if (world_size % 2)
{
TLLM_LOG_WARNING("world size is not a multiple of 2, return");
return;
}
int iter = 100;
std::vector<int> candidate_hidden_dim{1024, 2048, 4096, 7168, 8192};
int min_token_num = 1;
int max_token_num = 2048;
for (auto hidden_dim : candidate_hidden_dim)
{
Runner runner(max_token_num, hidden_dim);
for (int i = 0; i < iter; ++i)
{
int token_num = get_random_int(min_token_num, max_token_num);
if (rank == 0)
{
printf("[Verify] token_num %-4d, hidden_dim %-4d ...", token_num, hidden_dim);
}
runner.reset_io();
runner.run_once(&Runner::run_kernel, token_num, hidden_dim);
runner.verify(token_num, hidden_dim);
if (rank == 0)
{
printf("\033[32mPass!\033[0m\n");
}
}
}
}
TEST(Kernel_AllReduceFusion, AllReduceAccuracyFixedTokenNum)
{
using Runner = TestRunner<half, ar_fusion::AllReduceFusionPattern::kAllReduce>;
auto& comm = mpi::MpiComm::world();
auto world_size = comm.getSize();
auto rank = comm.getRank();
if (world_size % 2)
{
TLLM_LOG_WARNING("world size is not a multiple of 2, return");
return;
}
int iter = 10;
std::vector<int> candidate_hidden_dim{1024, 2048, 4096, 7168, 8192};
int min_token_num = 1;
int max_token_num = 2048;
for (auto hidden_dim : candidate_hidden_dim)
{
Runner runner(max_token_num, hidden_dim);
for (int token_num = min_token_num; token_num <= max_token_num; token_num *= 2)
{
if (rank == 0)
{
printf("[Verify] token_num %-4d, hidden_dim %-4d ...", token_num, hidden_dim);
}
for (int i = 0; i < iter; ++i)
{
runner.reset_io();
runner.run_once(&Runner::run_kernel, token_num, hidden_dim);
runner.verify(token_num, hidden_dim);
}
if (rank == 0)
{
printf("\033[32mPass!\033[0m\n");
}
}
}
}
TEST(Kernel_AllReduceFusion, AllReduceFusionAccuracyDifferentHiddenDim)
{
#define TEST_AR_FUSION(DType, FusionPattern) \
{ \
using Runner = TestRunner<DType, FusionPattern>; \
int iter = 10; \
std::vector<int> candidate_hidden_dim{64, 128, 256, 384, 512, 640, 768, 896}; \
int min_token_num = 1; \
int max_token_num = 2048; \
for (auto hidden_dim : candidate_hidden_dim) \
{ \
Runner runner(max_token_num, hidden_dim); \
for (int token_num = min_token_num; token_num <= max_token_num; token_num *= 2) \
{ \
if (rank == 0) \
{ \
printf("[Verify] token_num %-4d, hidden_dim %-4d ...", token_num, hidden_dim); \
} \
for (int i = 0; i < iter; ++i) \
{ \
runner.reset_io(); \
runner.run_once(&Runner::run_kernel, token_num, hidden_dim); \
runner.verify(token_num, hidden_dim); \
} \
if (rank == 0) \
{ \
printf("\033[32mPass!\033[0m\n"); \
} \
} \
} \
}
auto& comm = mpi::MpiComm::world();
auto world_size = comm.getSize();
auto rank = comm.getRank();
if (world_size % 2)
{
TLLM_LOG_WARNING("world size is not a multiple of 2, return");
return;
}
int const arch = tensorrt_llm::common::getSMVersion();
if (arch >= 100)
{
TEST_AR_FUSION(half, ar_fusion::AllReduceFusionPattern::kARResidualRMSNormOutFP4Quant);
}
else
{
TEST_AR_FUSION(half, ar_fusion::AllReduceFusionPattern::kARResidualRMSNormOutFP8Quant);
}
#undef TEST_AR_FUSION
}
TEST(Kernel_AllReduceFusion, AllReduceFusionAccuracyDifferentDType)
{
#define TEST_AR_FUSION(DType, FusionPattern) \
{ \
using Runner = TestRunner<DType, FusionPattern>; \
Runner runner(max_token_num, hidden_dim); \
for (int token_num = min_token_num; token_num <= max_token_num; token_num *= 2) \
{ \
if (rank == 0) \
{ \
printf("[Verify] pattern %-20s, dtype %-10s, token_num %-4d, hidden_dim %-4d ...", #FusionPattern, \
#DType, token_num, hidden_dim); \
} \
runner.reset_io(); \
runner.run_once(&Runner::run_kernel, token_num, hidden_dim); \
runner.verify(token_num, hidden_dim); \
if (rank == 0) \
{ \
printf("\033[32mPass!\033[0m\n"); \
} \
} \
}
int const arch = tensorrt_llm::common::getSMVersion();
auto& comm = mpi::MpiComm::world();
auto world_size = comm.getSize();
auto rank = comm.getRank();
if (world_size % 2)
{
TLLM_LOG_WARNING("world size is not a multiple of 2, return");
return;
}
std::vector<int> candidate_hidden_dim{1024, 2048, 4096, 7168, 8192};
int min_token_num = 1;
int max_token_num = 2048;
for (auto hidden_dim : candidate_hidden_dim)
{
TEST_AR_FUSION(half, ar_fusion::AllReduceFusionPattern::kAllReduce);
TEST_AR_FUSION(half, ar_fusion::AllReduceFusionPattern::kARResidualRMSNorm);
TEST_AR_FUSION(half, ar_fusion::AllReduceFusionPattern::kARResidualRMSNormOutFP8Quant);
if (arch >= 100)
{
TEST_AR_FUSION(half, ar_fusion::AllReduceFusionPattern::kARResidualRMSNormOutFP4Quant);
}
TEST_AR_FUSION(float, ar_fusion::AllReduceFusionPattern::kAllReduce);
TEST_AR_FUSION(float, ar_fusion::AllReduceFusionPattern::kARResidualRMSNorm);
TEST_AR_FUSION(float, ar_fusion::AllReduceFusionPattern::kARResidualRMSNormOutFP8Quant);
#if defined(ENABLE_BF16)
TEST_AR_FUSION(__nv_bfloat16, ar_fusion::AllReduceFusionPattern::kAllReduce);
TEST_AR_FUSION(__nv_bfloat16, ar_fusion::AllReduceFusionPattern::kARResidualRMSNorm);
TEST_AR_FUSION(__nv_bfloat16, ar_fusion::AllReduceFusionPattern::kARResidualRMSNormOutFP8Quant);
if (arch >= 100)
{
TEST_AR_FUSION(__nv_bfloat16, ar_fusion::AllReduceFusionPattern::kARResidualRMSNormOutFP4Quant);
}
#endif
}
#undef TEST_AR_FUSION
}
TEST(Kernel_AllReduceFusion, Perf)
{
int const arch = tensorrt_llm::common::getSMVersion();
if (arch >= 100)
{
using Runner = TestRunner<half, ar_fusion::AllReduceFusionPattern::kARResidualRMSNormFP4Quant>;
auto& comm = mpi::MpiComm::world();
auto world_size = comm.getSize();
auto rank = comm.getRank();
if (world_size % 2)
{
TLLM_LOG_WARNING("world size is not a multiple of 2, return");
return;
}
int warmup = 100, iter = 300;
int hidden_dim = 7168;
std::vector<int> candidate_token_num{1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1024, 2048};
int max_token_num = 2048;
Runner runner(max_token_num, hidden_dim);
for (auto token_num : candidate_token_num)
{
auto latency = runner.benchmark(&Runner::run_kernel, warmup, iter, token_num, hidden_dim);
if (rank == 0)
{
TLLM_LOG_INFO(
"token_num %-4d, hidden_dim %-4d, fusion kernel latency %4.4fus", token_num, hidden_dim, latency);
}
auto nccl_latency = runner.benchmark(&Runner::run_nccl_allreduce, warmup, iter, token_num, hidden_dim);
if (rank == 0)
{
TLLM_LOG_INFO("nccl allreduce latency %4.4fus", nccl_latency);
}
auto residual_latency = runner.benchmark(&Runner::run_residual_add, warmup, iter, token_num, hidden_dim);
if (rank == 0)
{
TLLM_LOG_INFO("residual add latency %4.4fus", residual_latency);
}
auto rms_latency = runner.benchmark(&Runner::run_rms_norm, warmup, iter, token_num, hidden_dim);
if (rank == 0)
{
TLLM_LOG_INFO("rms norm latency %4.4fus", rms_latency);
}
auto quant_latency = runner.benchmark(&Runner::run_fp4_quant, warmup, iter, token_num, hidden_dim);
if (rank == 0)
{
TLLM_LOG_INFO("fp4 quant latency %4.4fus", quant_latency);
auto tot_latency = nccl_latency + residual_latency + rms_latency + quant_latency;
TLLM_LOG_INFO("fusion kernel latency %4.4fus, nccl + ops latency %4.4fus, total speedup %2.4fx",
latency, tot_latency, tot_latency / latency);
}
}
}
}