mirror of
https://github.com/pytorch/pytorch.git
synced 2025-10-20 12:54:11 +08:00
[ATen][CMake] Fix duplicated CUTLASS path (#165424)
Fixes #165110 The `PUBLIC` scope causes CUTLASS of the FBGEMM being included in for all PyTorch targets, including special matmuls (RowwiseScaledMM, ScaledGroupMM and GroupMM). Due to version mismatch between FBGEMM/CUTLASS and PyTorch/CUTLASS it is unacceptable to use FBGEMM/CUTLASS in PyTorch targets. This PR limits the scope of FBGEMM/CUTLASS to `fbgemm_genai` target only. Pull Request resolved: https://github.com/pytorch/pytorch/pull/165424 Approved by: https://github.com/cthi, https://github.com/eqy, https://github.com/danielvegamyhre
This commit is contained in:
Aidyn-A
committed by
PyTorch MergeBot
PyTorch MergeBot
parent
712f54d453
commit
7719cb75bf
@ -256,6 +256,7 @@ endif()
|
||||
IF(USE_FBGEMM_GENAI)
|
||||
set(FBGEMM_THIRD_PARTY ${PROJECT_SOURCE_DIR}/third_party/fbgemm/external/)
|
||||
set(FBGEMM_GENAI_SRCS ${PROJECT_SOURCE_DIR}/third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize)
|
||||
|
||||
if(USE_CUDA)
|
||||
# To avoid increasing the build time/binary size unnecessarily, use an allow-list of kernels to build.
|
||||
# If you want to integrate a kernel from FBGEMM into torch, you have to add it here.
|
||||
@ -292,58 +293,64 @@ IF(USE_FBGEMM_GENAI)
|
||||
"${FBGEMM_GENAI_SRCS}/cutlass_extensions/mx8mx8bf16_grouped/"
|
||||
)
|
||||
|
||||
target_include_directories(fbgemm_genai PUBLIC
|
||||
target_include_directories(fbgemm_genai PRIVATE
|
||||
${FBGEMM_THIRD_PARTY}/cutlass/include
|
||||
${FBGEMM_THIRD_PARTY}/cutlass/tools/util/include
|
||||
${fbgemm_genai_mx8mx8bf16_grouped}
|
||||
${FBGEMM_GENAI_SRCS}/common/include/ # includes fbgemm_gpu/quantize/utils.h, fbgemm_gpu/quantize/tuning_cache.hpp
|
||||
${FBGEMM_GENAI_SRCS}/include/ # includes fbgemm_gpu/torch_ops.h
|
||||
)
|
||||
else()
|
||||
if(USE_ROCM)
|
||||
# Only include the kernels we want to build to avoid increasing binary size.
|
||||
file(GLOB_RECURSE fbgemm_genai_native_rocm_hip
|
||||
"${FBGEMM_GENAI_SRCS}/ck_extensions/fp8_rowwise_grouped/kernels/fp8_rowwise_grouped*.hip"
|
||||
"${FBGEMM_GENAI_SRCS}/ck_extensions/fp8_rowwise_grouped/fp8_rowwise_grouped_gemm.hip")
|
||||
set_source_files_properties(${fbgemm_genai_native_rocm_hip} PROPERTIES HIP_SOURCE_PROPERTY_FORMAT 1)
|
||||
|
||||
# Add additional HIPCC compiler flags for performance
|
||||
set(FBGEMM_GENAI_EXTRA_HIPCC_FLAGS
|
||||
-mllvm
|
||||
-amdgpu-coerce-illegal-types=1
|
||||
-mllvm
|
||||
-enable-post-misched=0
|
||||
-mllvm
|
||||
-greedy-reverse-local-assignment=1
|
||||
-fhip-new-launch-api)
|
||||
# Add FBGEMM_GENAI include directories for torch_ops.h
|
||||
list(APPEND ATen_CUDA_INCLUDE ${PROJECT_SOURCE_DIR}/third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/include)
|
||||
list(APPEND ATen_CUDA_INCLUDE ${PROJECT_SOURCE_DIR}/third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/common/include)
|
||||
elseif(USE_ROCM)
|
||||
# Only include the kernels we want to build to avoid increasing binary size.
|
||||
file(GLOB_RECURSE fbgemm_genai_native_rocm_hip
|
||||
"${FBGEMM_GENAI_SRCS}/ck_extensions/fp8_rowwise_grouped/kernels/fp8_rowwise_grouped*.hip"
|
||||
"${FBGEMM_GENAI_SRCS}/ck_extensions/fp8_rowwise_grouped/fp8_rowwise_grouped_gemm.hip")
|
||||
set_source_files_properties(${fbgemm_genai_native_rocm_hip} PROPERTIES HIP_SOURCE_PROPERTY_FORMAT 1)
|
||||
|
||||
# Only compile for gfx942 for now.
|
||||
# This is rather hacky, I could not figure out a clean solution :(
|
||||
set(HIP_CLANG_FLAGS_ORIGINAL ${HIP_CLANG_FLAGS})
|
||||
string(REGEX REPLACE "--offload-arch=[^ ]*" "" FILTERED_HIP_CLANG_FLAGS "${HIP_CLANG_FLAGS}")
|
||||
if("gfx942" IN_LIST PYTORCH_ROCM_ARCH)
|
||||
list(APPEND FILTERED_HIP_CLANG_FLAGS --offload-arch=gfx942;)
|
||||
endif()
|
||||
set(HIP_CLANG_FLAGS ${FILTERED_HIP_CLANG_FLAGS})
|
||||
# Add additional HIPCC compiler flags for performance
|
||||
set(FBGEMM_GENAI_EXTRA_HIPCC_FLAGS
|
||||
-mllvm
|
||||
-amdgpu-coerce-illegal-types=1
|
||||
-mllvm
|
||||
-enable-post-misched=0
|
||||
-mllvm
|
||||
-greedy-reverse-local-assignment=1
|
||||
-fhip-new-launch-api)
|
||||
|
||||
hip_add_library(
|
||||
fbgemm_genai STATIC
|
||||
${fbgemm_genai_native_rocm_hip}
|
||||
HIPCC_OPTIONS ${HIP_HCC_FLAGS} ${FBGEMM_GENAI_EXTRA_HIPCC_FLAGS})
|
||||
set(HIP_CLANG_FLAGS ${HIP_CLANG_FLAGS_ORIGINAL})
|
||||
set_target_properties(fbgemm_genai PROPERTIES POSITION_INDEPENDENT_CODE ON)
|
||||
target_compile_definitions(fbgemm_genai PRIVATE FBGEMM_GENAI_NO_EXTENDED_SHAPES)
|
||||
|
||||
target_include_directories(fbgemm_genai PUBLIC
|
||||
# FBGEMM version of Composable Kernel is used due to some customizations
|
||||
${FBGEMM_THIRD_PARTY}/composable_kernel/include
|
||||
${FBGEMM_THIRD_PARTY}/composable_kernel/library/include
|
||||
${FBGEMM_THIRD_PARTY}/cutlass/include
|
||||
${FBGEMM_THIRD_PARTY}/cutlass/tools/util/include
|
||||
${FBGEMM_GENAI_SRCS}/common/include/ # includes fbgemm_gpu/quantize/utils.h, fbgemm_gpu/quantize/tuning_cache.hpp
|
||||
${FBGEMM_GENAI_SRCS}/include/ # includes fbgemm_gpu/torch_ops.h
|
||||
)
|
||||
# Only compile for gfx942 for now.
|
||||
# This is rather hacky, I could not figure out a clean solution :(
|
||||
set(HIP_CLANG_FLAGS_ORIGINAL ${HIP_CLANG_FLAGS})
|
||||
string(REGEX REPLACE "--offload-arch=[^ ]*" "" FILTERED_HIP_CLANG_FLAGS "${HIP_CLANG_FLAGS}")
|
||||
if("gfx942" IN_LIST PYTORCH_ROCM_ARCH)
|
||||
list(APPEND FILTERED_HIP_CLANG_FLAGS --offload-arch=gfx942;)
|
||||
endif()
|
||||
set(HIP_CLANG_FLAGS ${FILTERED_HIP_CLANG_FLAGS})
|
||||
|
||||
hip_add_library(
|
||||
fbgemm_genai STATIC
|
||||
${fbgemm_genai_native_rocm_hip}
|
||||
HIPCC_OPTIONS ${HIP_HCC_FLAGS} ${FBGEMM_GENAI_EXTRA_HIPCC_FLAGS})
|
||||
set(HIP_CLANG_FLAGS ${HIP_CLANG_FLAGS_ORIGINAL})
|
||||
set_target_properties(fbgemm_genai PROPERTIES POSITION_INDEPENDENT_CODE ON)
|
||||
target_compile_definitions(fbgemm_genai PRIVATE FBGEMM_GENAI_NO_EXTENDED_SHAPES)
|
||||
|
||||
target_include_directories(fbgemm_genai PRIVATE
|
||||
# FBGEMM version of Composable Kernel is used due to some customizations
|
||||
${FBGEMM_THIRD_PARTY}/composable_kernel/include
|
||||
${FBGEMM_THIRD_PARTY}/composable_kernel/library/include
|
||||
${FBGEMM_THIRD_PARTY}/cutlass/include
|
||||
${FBGEMM_THIRD_PARTY}/cutlass/tools/util/include
|
||||
${FBGEMM_GENAI_SRCS}/common/include/ # includes fbgemm_gpu/quantize/utils.h, fbgemm_gpu/quantize/tuning_cache.hpp
|
||||
${FBGEMM_GENAI_SRCS}/include/ # includes fbgemm_gpu/torch_ops.h
|
||||
)
|
||||
|
||||
# Add FBGEMM_GENAI include directories for torch_ops.h
|
||||
list(APPEND ATen_HIP_INCLUDE ${PROJECT_SOURCE_DIR}/third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/include)
|
||||
list(APPEND ATen_HIP_INCLUDE ${PROJECT_SOURCE_DIR}/third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/common/include)
|
||||
endif()
|
||||
endif()
|
||||
|
||||
@ -692,12 +699,6 @@ if(USE_CUDA AND NOT USE_ROCM)
|
||||
list(APPEND ATen_CUDA_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/cutlass/include)
|
||||
list(APPEND ATen_CUDA_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/cutlass/tools/util/include)
|
||||
|
||||
# Add FBGEMM_GENAI include directories for torch_ops.h
|
||||
if(USE_FBGEMM_GENAI)
|
||||
list(APPEND ATen_CUDA_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/include)
|
||||
list(APPEND ATen_CUDA_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/common/include)
|
||||
endif()
|
||||
|
||||
if($ENV{ATEN_STATIC_CUDA})
|
||||
if(CUDA_VERSION VERSION_LESS_EQUAL 12.9)
|
||||
list(APPEND ATen_CUDA_DEPENDENCY_LIBS
|
||||
|
||||
Reference in New Issue
Block a user