Build Instrctions of Pytorch 1.0.1 on Mac OS X 10.13

By TomHeaven @ 2019.4.3

If you only want to use pytorch on Mac with CUDA support, first go to release page: https://github.com/TomHeaven/pytorch-osx-build/releases and try to find proper wheel pakcages.

However, as their are mulitple versions of python (2.7, 3.6, 3.7) and CUDA (8.0, 9.0, 9.2, 10.0). It's more and more time consuming to support all combinations of them. If your combination of python and CUDA is not listed in the release page, you may need to compile a wheel package yourself.

This guide will help you compile a wheel package of pytorch-1.0.1 on Mac OS X 10.13 with CUDA support.

Dependencies

Dependencies:

Mac OS 10.13.x.
XCode 9.4.1
CUDA 10.0 and CuDNN7.4
Optinally, Openblas for CPU accelerated matrix computation and so on. In previous releases (about 0.3.0 or 0.4.0), Pytorch cannot correctly detect vecLib on Mac OS. So we need Openblas as an alternative. However, this problem seems to be solved now and we no longer have to install openblas.

brew install openblas
brew link --overwrite openblas

Optinally, OpenMP for multi-thread computation. Install using Homebrew:

brew install libomp
brew link --overwrite libomp

【Most of you don't need this.】Optionally, OpenMPI for distributed computation. The default supported backend on Mac OS is tcp. With OpenMPI, you can use mpi as backend in torch.distributed.deprecated (The newly constructed torch.distributed does not support Mac OS yet.) and enjoy distributed training with multiple machines and GPUs.

brew install openmpi
brew link --overwrite openmpi

Distributed computation sounds great but requires extra efforts in coding. See https://pytorch.org/tutorials/intermediate/dist_tuto.html.

I'm also eagerly waiting for NCCL, a better distributed backend, for Mac OS and opened an issue NVIDIA/nccl#201 there to make developers in Nvidia aware of the request.

Clone Pytorch Repo

Install git and clone pytorch repo by

brew install git
git clone https://github.com/pytorch/pytorch
cd pytorch
git checkout v1.0.1
git submodule update --init

Update Dependency.cmake

If you do not need OpenMP or MPI for distributed computation, just skip this section and start compiling.

The default OpenMP detection will fail. If you need OpenMP, manually set related build flags by

modifying cmake/Dependency.cmake:448 as

# ---[ OpenMP
#if(USE_OPENMP)
  #find_package(OpenMP)
  
  set(OpenMP_C_FLAGS "-Xpreprocessor -fopenmp -I/usr/local/include -L/usr/local/lib -lomp -lgomp")
  set(OpenMP_CXX_FLAGS "-Xpreprocessor -fopenmp")
  set(OpenMP_EXE_LINKER_FLAGS "-Xpreprocessor -fopenmp -L/usr/local/lib -lomp -lgomp")
  caffe2_update_option(USE_OPENMP ON)

  #if(OPENMP_FOUND)
    message(STATUS "Adding " ${OpenMP_CXX_FLAGS})
    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${OpenMP_C_FLAGS}")
    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${OpenMP_CXX_FLAGS}")
    set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} ${OpenMP_EXE_LINKER_FLAGS}")
  #else()
  #  message(WARNING "Not compiling with OpenMP. Suppress this warning with -DUSE_OPENMP=OFF")
   # caffe2_update_option(USE_OPENMP OFF)
  #endif()
#endif()

to make caffe2 use OpenMP.

modify third_party/ideep/mkl-dnn/cmake/OpenMP.cmake:93 as

#set(OpenMP_CXX_FLAGS "-fopenmp")
#set(OpenMP_C_FLAGS "-fopenmp")
set(OpenMP_C_FLAGS "-Xpreprocessor -fopenmp -I/usr/local/include -L/usr/local/lib -lomp -lgomp")
set(OpenMP_CXX_FLAGS "-Xpreprocessor -fopenmp -I/usr/local/include -L/usr/local/lib -lomp -lgomp")

to avoid a -fopenmp related link error at about 90%.

modify torch/CMakeListst.txt:330 as

find_package(OpenMP)
if(OPENMP_FOUND)
  if (VERBOSE)
    message(STATUS "Compiling with OpenMP")
  endif()
  target_compile_options(torch INTERFACE -Xpreprocessor -fopenmp)
  target_link_libraries(torch -Xpreprocessor -fopenmp)
endif()

to avoid a -fopenmp related link error at about 95%.

Compilation

Compile using

# version
export PYTORCH_BUILD_VERSION=1.0.1
export PYTORCH_BUILD_NUMBER=1
# set CUDA_ARCH
export TORCH_CUDA_ARCH_LIST="3.5;5.2;6.1;7.0"
# compile for python2
pip install pyyaml
python setup.py bdist_wheel
# compile for python3
pip3 install pyyaml
python3 setup.py bdist_wheel

The wheel packages will be avaiable at dist folder if everything goes smoothly.

Installation

Just cd to dist and install wheel packages using pip.

cd dist
pip install pytorch-xxx-cp27-xxx.wheel
pip3 install pytorch-xxx-cp3x-xxx.wheel