mirror of
https://github.com/NVIDIA/TensorRT-LLM.git
synced 2026-01-29 23:23:48 +08:00
16d2467ea8
* Update TensorRT-LLM --------- Co-authored-by: Denis Kayshev <topenkoff@gmail.com> Co-authored-by: akhoroshev <arthoroshev@gmail.com> Co-authored-by: Patrick Reiter Horn <patrick.horn@gmail.com> Update
80 lines
3.4 KiB
CMake
80 lines
3.4 KiB
CMake
# SPDX-FileCopyrightText: Copyright (c) 2023-2024 NVIDIA CORPORATION &
|
|
# AFFILIATES. All rights reserved. SPDX-License-Identifier: NVIDIA TensorRT
|
|
# Source Code License Agreement
|
|
#
|
|
# NVIDIA CORPORATION, its affiliates and licensors retain all intellectual
|
|
# property and proprietary rights in and to this material, related documentation
|
|
# and any modifications thereto. Any use, reproduction, disclosure or
|
|
# distribution of this material and related documentation without an express
|
|
# license agreement from NVIDIA CORPORATION or its affiliates is strictly
|
|
# prohibited.
|
|
|
|
add_gtest(banRepeatNGramsKernelsTest banRepeatNGramsKernelsTest.cpp)
|
|
add_gtest(decodingKernelsTest decodingKernelTest.cpp)
|
|
add_gtest(logitsBitmaskTest logitsBitmaskTest.cpp)
|
|
add_gtest(mixtureOfExpertsTest mixtureOfExpertsTest.cu)
|
|
add_gtest(ropeTest ropeTest.cu)
|
|
add_gtest(shiftKCacheKernelTest shiftKCacheKernelTest.cu)
|
|
add_gtest(smoothQuantKernelTest smoothQuant/smoothQuantKernelTest.cpp)
|
|
add_gtest(stopCriteriaKernelsTest stopCriteriaKernelsTest.cpp)
|
|
add_gtest(weightOnlyKernelTest weightOnly/weightOnlyKernelTest.cpp)
|
|
|
|
add_gtest(cudaCoreGemmKernelTest cudaCoreGemm/cudaCoreGemmKernelTest.cpp)
|
|
|
|
if(NOT ENABLE_MULTI_DEVICE EQUAL 0)
|
|
add_gtest(allReduceKernelTest allReduce/allReduceKernelTest.cu)
|
|
add_gtest(gemmAllReduceTest allReduce/gemmAllReduceTest.cu)
|
|
endif()
|
|
|
|
add_gtest(
|
|
gemmSwigluRunnerTest
|
|
fused_gated_gemm/gemmSwigluRunnerTest.cu
|
|
${PROJECT_SOURCE_DIR}/tensorrt_llm/cutlass_extensions/kernels/fused_gated_gemm/gemm_swiglu_e4m3.cu
|
|
NO_GTEST_MAIN)
|
|
add_gtest(gemmSwigluKernelTestSm90Fp8
|
|
fused_gated_gemm/gemmSwigluKernelTestSm90Fp8.cu NO_GTEST_MAIN
|
|
NO_TLLM_LINKAGE)
|
|
|
|
foreach(target_name gemmSwigluRunnerTest;gemmSwigluKernelTestSm90Fp8)
|
|
set_property(TARGET ${target_name} PROPERTY CUDA_RESOLVE_DEVICE_SYMBOLS ON)
|
|
|
|
# Note - we deliberately do not include 90a PTX (even when 9.0+PTX is
|
|
# specified). This is because sm_90a has arch conditional instructions that
|
|
# are not forward compatible. As a result, it does not make sense to embed PTX
|
|
# into the binary anyway.
|
|
if("90" IN_LIST CMAKE_CUDA_ARCHITECTURES_ORIG
|
|
OR "90-real" IN_LIST CMAKE_CUDA_ARCHITECTURES_ORIG
|
|
OR "90-real" IN_LIST CMAKE_CUDA_ARCHITECTURES_NATIVE)
|
|
|
|
message(STATUS "MANUALLY APPENDING FLAG TO COMPILE FOR SM_90a.")
|
|
target_compile_options(
|
|
${target_name}
|
|
PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:-gencode=arch=compute_90a,code=sm_90a
|
|
-res-usage>)
|
|
|
|
# Hopper kernels require cuda lib for TMA APIs
|
|
target_link_libraries(${target_name} PRIVATE CUDA::cuda_driver)
|
|
|
|
# No kernels should be parsed, unless hopper is specified. This is a build
|
|
# time improvement
|
|
target_compile_definitions(${target_name} PRIVATE COMPILE_HOPPER_TMA_GEMMS)
|
|
target_compile_definitions(${target_name}
|
|
PRIVATE COMPILE_HOPPER_TMA_GROUPED_GEMMS)
|
|
endif()
|
|
|
|
# Suppress GCC note: the ABI for passing parameters with 64-byte alignment has
|
|
# changed in GCC 4.6 This note appears for kernels using TMA and clutters the
|
|
# compilation output.
|
|
if(NOT WIN32)
|
|
target_compile_options(
|
|
${target_name} PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:-Xcompiler=-Wno-psabi>)
|
|
endif()
|
|
endforeach()
|
|
|
|
set(SAMPLING_KERNEL_TEST_SRC
|
|
sampling/samplingTest.cpp sampling/samplingTopKTest.cpp
|
|
sampling/samplingTopPTest.cpp sampling/samplingAirTopPTest.cpp
|
|
sampling/samplingPenaltyTest.cpp sampling/samplingUtilsTest.cu)
|
|
|
|
add_gtest(samplingKernelsTest "${SAMPLING_KERNEL_TEST_SRC}")
|