OSSForks/pytorch
1
0
Fork 0
mirror of https://github.com/zebrajr/pytorch.git synced 2025-12-06 12:20:52 +01:00
Code Issues Actions 38 Packages Projects Releases Wiki Activity

Add torch_cpu specific flag for debug info (#57190)

Browse Source 
Summary:
Right now we are using `REL_WITH_DEB_INFO=1` on Linux CI binary builds. This is causing intermittent failures on CUDA builds since the debug information increases the load on the linker. This adds a workaround by a flag to enable debug info only for the target we actually want it for (`libtorch_cpu.so`, all the other binaries are stripped over their debug info after building).

Example failures (from [the hud](https://ezyang.github.io/pytorch-ci-hud/build2/pytorch-nightly?mode=nightly)):
* https://app.circleci.com/pipelines/github/pytorch/pytorch/311785/workflows/df640957-54b0-4592-aeef-6d5baee503ae/jobs/12932229
* https://app.circleci.com/pipelines/github/pytorch/pytorch/311784/workflows/e3b487d6-fb46-4a5d-a2d5-22eec328b678/jobs/12932228
* https://app.circleci.com/pipelines/github/pytorch/pytorch/311784/workflows/e3b487d6-fb46-4a5d-a2d5-22eec328b678/jobs/12932227

Pull Request resolved: https://github.com/pytorch/pytorch/pull/57190

Pulled By: driazati

Reviewed By: janeyx99

Differential Revision: D28085550

fbshipit-source-id: 0fc5b3e769b10c0dd3811717f968d0c933667361
This commit is contained in:
davidriazati@fb.com 2021-04-29 12:02:51 -07:00 committed by Facebook GitHub Bot
parent d3ffe9ab6b
commit 21be40b390
3 changed files with 12 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
.circleci/scripts/binary_checkout.sh
View File

@ -64,5 +64,7 @@ popd
retry git clone -q https://github.com/pytorch/builder.git "$BUILDER_ROOT" retry git clone -q https://github.com/pytorch/builder.git "$BUILDER_ROOT"
pushd "$BUILDER_ROOT" pushd "$BUILDER_ROOT"
echo "Using builder from " echo "Using builder from "
# TODO: Remove before landing, this is just for testing
git checkout driazati/torch_debug_flag
git --no-pager log --max-count 1 git --no-pager log --max-count 1
popd popd

1
CMakeLists.txt
View File

@ -282,6 +282,7 @@ cmake_dependent_option(
option(USE_TBB "Use TBB" OFF) option(USE_TBB "Use TBB" OFF)
option(ONNX_ML "Enable traditional ONNX ML API." ON) option(ONNX_ML "Enable traditional ONNX ML API." ON)
option(HAVE_SOVERSION "Whether to add SOVERSION to the shared objects" OFF) option(HAVE_SOVERSION "Whether to add SOVERSION to the shared objects" OFF)
option(BUILD_LIBTORCH_CPU_WITH_DEBUG "Enable RelWithDebInfo for libtorch_cpu target only" OFF)
cmake_dependent_option( cmake_dependent_option(
USE_DEPLOY "Build embedded torch::deploy interpreter. See torch/csrc/deploy/README.md for more info." OFF USE_DEPLOY "Build embedded torch::deploy interpreter. See torch/csrc/deploy/README.md for more info." OFF
"BUILD_PYTHON" OFF) "BUILD_PYTHON" OFF)

9
caffe2/CMakeLists.txt
View File

@ -770,6 +770,15 @@ if(HAVE_SOVERSION)
endif() endif()
torch_compile_options(torch_cpu) # see cmake/public/utils.cmake torch_compile_options(torch_cpu) # see cmake/public/utils.cmake
if(CMAKE_COMPILER_IS_GNUCXX AND BUILD_LIBTORCH_CPU_WITH_DEBUG)
# To enable debug fission we need to build libtorch_cpu with debug info on,
# but this increases link time and peak memory usage if we use the
# REL_WITH_DEB_INFO env var since that enables it for everything, but it's
# only really necessary for libtorch_cpu.
target_compile_options(torch_cpu PRIVATE "-g")
endif()
if(USE_LLVM AND LLVM_FOUND) if(USE_LLVM AND LLVM_FOUND)
llvm_map_components_to_libnames(LLVM_LINK_LIBS llvm_map_components_to_libnames(LLVM_LINK_LIBS
support core analysis executionengine instcombine support core analysis executionengine instcombine