Anyone working on binary distribution for Mac w/ Cuda?

1

Is there any hope that there will be a binary release of Pytorch with Cuda for Mac OS X soon? I know that it can be compiled from source but with complications such as downgrading to an earlier version of xtools, manually copying header files, etc, but really don’t want to go that route! I’m sure that many other Mac users feel the same…

2

I’m interested in this as well. I’ve been trying to compile from source, but am having issues. The xtools version issue makes sense.

A binary package with Cuda support would be really nice.

Edit: I was finally able to build PyTorch from source (using the Xcode 9 toolchain - it’s all I have installed). My issue was due to environment variables and fish shell. I instead used bash and the typical MACOSX_DEPLOYMENT_TARGET=10.9 CC=clang CXX=clang++ python setup.py install everyone suggests and it appears to have worked.

3

I followed this guide:

This hint is shown on the from page pytorch.org e.g. when
[OS] OSX
[Package Manager] Source
[Python] 3.6
[CUDA] 9.1

The benefit is if you have conda installed there is no need to deal with all the X Tools stuff, you can simply use cmake from conda for compilation.

Citation from the guide mentioned:

export CMAKE_PREFIX_PATH=[anaconda root directory]
conda install numpy pyyaml setuptools cmake cffi typing

git clone --recursive https://github.com/pytorch/pytorch
cd pytorch

# fix OS Version if nor on 10.13
MACOSX_DEPLOYMENT_TARGET=10.13 CC=clang CXX=clang++ python setup.py install
1 Like
4

I am trying this as well, and in regard for the CLANG version, this should help;