frozenleaves/pytorch
1
0
Fork 0
mirror of https://github.com/pytorch/pytorch.git synced 2025-10-20 21:14:14 +08:00
Code Issues Packages Projects Releases Wiki Activity

[BE] Delete NCCL slimming (#146943)

Browse Source 
It was added by https://github.com/pytorch/pytorch/pull/35843 and served its purpose when everything was linked statically in libtorch_cuda.so, but for all our releases it's no longer relevant as nccl is now a dynamic dependency of libtorch_cuda.so

Besides,  It does not work with CXX11 ABI anyway, and creates problems with newer version of NCCL, when two `collectvies.o` are package into library archive.

Pull Request resolved: https://github.com/pytorch/pytorch/pull/146943
Approved by: https://github.com/Skylion007, https://github.com/atalman
This commit is contained in:
Nikita Shulga
2025-02-12 00:35:55 +00:00
committed by PyTorch MergeBot
parent a58f421f4b
commit df5e232563
1 changed files with 2 additions and 36 deletions
Download Patch File Download Diff File Expand all files Collapse all files

38
cmake/External/nccl.cmake vendored
View File

@ -55,42 +55,8 @@ if(NOT __NCCL_INCLUDED)
INSTALL_COMMAND ""
)
# Detect objcopy version
execute_process(COMMAND "${CMAKE_OBJCOPY}" "--version" OUTPUT_VARIABLE OBJCOPY_VERSION_STR)
string(REGEX REPLACE "GNU objcopy .+ ([0-9])\\.([0-9]+).*" "\\1" OBJCOPY_VERSION_MAJOR ${OBJCOPY_VERSION_STR})
string(REGEX REPLACE "GNU objcopy .+ ([0-9])\\.([0-9]+).*" "\\2" OBJCOPY_VERSION_MINOR ${OBJCOPY_VERSION_STR})
# TODO: Replace me with SKIP_NCCL_SLIMMING option (and investigate why it does not work on newer compilers)
if("$ENV{BUILD_ENVIRONMENT}" MATCHES ".*-libtorch-cxx11-abi$")
# See https://github.com/pytorch/pytorch/issues/83887
message(WARNING "Skip NCCL library slimming for cxx11-abi builds")
set(__NCCL_LIBRARY_DEP nccl_external)
set(NCCL_LIBRARIES ${__NCCL_BUILD_DIR}/lib/libnccl_static.a)
elseif((${OBJCOPY_VERSION_MAJOR} GREATER 2) OR ((${OBJCOPY_VERSION_MAJOR} EQUAL 2) AND (${OBJCOPY_VERSION_MINOR} GREATER 27)))
message(WARNING "Enabling NCCL library slimming")
add_custom_command(
OUTPUT "${__NCCL_BUILD_DIR}/lib/libnccl_slim_static.a"
DEPENDS nccl_external
COMMAND "${CMAKE_COMMAND}" -E make_directory "${__NCCL_BUILD_DIR}/objects"
COMMAND cd objects
COMMAND "${CMAKE_AR}" x "${__NCCL_BUILD_DIR}/lib/libnccl_static.a"
COMMAND for obj in all_gather_* all_reduce_* broadcast_* reduce_*.o$<SEMICOLON> do "${CMAKE_OBJCOPY}" --remove-relocations .nvFatBinSegment --remove-section __nv_relfatbin $$obj$<SEMICOLON> done
COMMAND "${CMAKE_AR}" cr "${__NCCL_BUILD_DIR}/lib/libnccl_slim_static.a" "*.o"
COMMAND "${CMAKE_AR}" xN 1 "${__NCCL_BUILD_DIR}/lib/libnccl_static.a" net.o
COMMAND "${CMAKE_AR}" q "${__NCCL_BUILD_DIR}/lib/libnccl_slim_static.a" net.o
COMMAND cd -
COMMAND "${CMAKE_COMMAND}" -E remove_directory "${__NCCL_BUILD_DIR}/objects"
WORKING_DIRECTORY "${__NCCL_BUILD_DIR}"
COMMENT "Slimming NCCL"
)
add_custom_target(nccl_slim_external DEPENDS "${__NCCL_BUILD_DIR}/lib/libnccl_slim_static.a")
set(__NCCL_LIBRARY_DEP nccl_slim_external)
set(NCCL_LIBRARIES ${__NCCL_BUILD_DIR}/lib/libnccl_slim_static.a)
else()
message(WARNING "Objcopy version is too old to support NCCL library slimming")
set(__NCCL_LIBRARY_DEP nccl_external)
set(NCCL_LIBRARIES ${__NCCL_BUILD_DIR}/lib/libnccl_static.a)
endif()
set(__NCCL_LIBRARY_DEP nccl_external)
set(NCCL_LIBRARIES ${__NCCL_BUILD_DIR}/lib/libnccl_static.a)
set(NCCL_FOUND TRUE)
add_library(__caffe2_nccl INTERFACE)