[Kernel][Misc] Use TORCH_LIBRARY instead of PYBIND11_MODULE for custom ops (#5047)

This commit is contained in:
bnellnm
2024-06-09 16:23:30 -04:00
committed by GitHub
parent 5d7e3d0176
commit 5467ac3196
55 changed files with 833 additions and 451 deletions

View File

@@ -12,7 +12,7 @@ include_directories("${CMAKE_SOURCE_DIR}/csrc")
#
# Check the compile flags
#
list(APPEND CXX_COMPILE_FLAGS
list(APPEND CXX_COMPILE_FLAGS
"-fopenmp"
"-DVLLM_CPU_EXTENSION")
@@ -44,8 +44,8 @@ if (AVX512_FOUND)
find_isa(${CPUINFO} "avx512_bf16" AVX512BF16_FOUND)
if (AVX512BF16_FOUND OR ENABLE_AVX512BF16)
if (CMAKE_CXX_COMPILER_ID STREQUAL "GNU" AND
CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 12.3)
if (CMAKE_CXX_COMPILER_ID STREQUAL "GNU" AND
CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 12.3)
list(APPEND CXX_COMPILE_FLAGS "-mavx512bf16")
else()
message(WARNING "Disable AVX512-BF16 ISA support, requires gcc/g++ >= 12.3")
@@ -73,7 +73,7 @@ set(VLLM_EXT_SRC
"csrc/cpu/cache.cpp"
"csrc/cpu/layernorm.cpp"
"csrc/cpu/pos_encoding.cpp"
"csrc/cpu/pybind.cpp")
"csrc/cpu/torch_bindings.cpp")
define_gpu_extension_target(
_C
@@ -81,10 +81,10 @@ define_gpu_extension_target(
LANGUAGE CXX
SOURCES ${VLLM_EXT_SRC}
COMPILE_FLAGS ${CXX_COMPILE_FLAGS}
WITH_SOABI
USE_SABI 3
WITH_SOABI
)
add_custom_target(default)
message(STATUS "Enabling C extension.")
add_dependencies(default _C)

View File

@@ -5,7 +5,7 @@
macro (find_python_from_executable EXECUTABLE SUPPORTED_VERSIONS)
file(REAL_PATH ${EXECUTABLE} EXECUTABLE)
set(Python_EXECUTABLE ${EXECUTABLE})
find_package(Python COMPONENTS Interpreter Development.Module)
find_package(Python COMPONENTS Interpreter Development.Module Development.SABIModule)
if (NOT Python_FOUND)
message(FATAL_ERROR "Unable to find python matching: ${EXECUTABLE}.")
endif()
@@ -294,6 +294,7 @@ endmacro()
# INCLUDE_DIRECTORIES <dirs> - Extra include directories.
# LIBRARIES <libraries> - Extra link libraries.
# WITH_SOABI - Generate library with python SOABI suffix name.
# USE_SABI <version> - Use python stable api <version>
#
# Note: optimization level/debug info is set via cmake build type.
#
@@ -301,7 +302,7 @@ function (define_gpu_extension_target GPU_MOD_NAME)
cmake_parse_arguments(PARSE_ARGV 1
GPU
"WITH_SOABI"
"DESTINATION;LANGUAGE"
"DESTINATION;LANGUAGE;USE_SABI"
"SOURCES;ARCHITECTURES;COMPILE_FLAGS;INCLUDE_DIRECTORIES;LIBRARIES")
# Add hipify preprocessing step when building with HIP/ROCm.
@@ -315,7 +316,11 @@ function (define_gpu_extension_target GPU_MOD_NAME)
set(GPU_WITH_SOABI)
endif()
Python_add_library(${GPU_MOD_NAME} MODULE "${GPU_SOURCES}" ${GPU_WITH_SOABI})
if (GPU_USE_SABI)
Python_add_library(${GPU_MOD_NAME} MODULE USE_SABI ${GPU_USE_SABI} ${GPU_WITH_SOABI} "${GPU_SOURCES}")
else()
Python_add_library(${GPU_MOD_NAME} MODULE ${GPU_WITH_SOABI} "${GPU_SOURCES}")
endif()
if (GPU_LANGUAGE STREQUAL "HIP")
# Make this target dependent on the hipify preprocessor step.