GPU driver and CUDA is not enabled and accessible by PyTorch on MacOS - macos

Question

GPU driver and CUDA is not enabled and accessible by PyTorch.
torch.cuda.is_available() returns false
I am using macOS Mojave 10.14.6
I have installed Cuda 10.0 version of pytorch.
I tried verfication on https://pytorch.org/get-started/locally/ and constructing a randomly initialized tensor works just fine.
But when I tried
import torch
torch.cuda.is_availalbe()
it returns false.
Therefore, I followed instructions on Pytorch and installed Anaconda and Cuda.
Then tried this:
git clone --recursive https://github.com/pytorch/pytorch
cd pytorch
MACOSX_DEPLOYMENT_TARGET=10.9 CC=clang CXX=clang++ python setup.py install
In terminal, I got
fatal error: 'string.h' file not found
#include_next <string.h>
I searched on stackoverflow and came up with this. Build Pytorch from source . So I tried:
$ find /Library/Developer/CommandLineTools/usr -type f -name string.h
which returned /Library/Developer/CommandLineTools/usr/include/c++/v1/string.h
Doesn't this mean I already have string.h?
How can I solve this problem?

Answer 1

Are you installing from a conda env? According to the github this should work:
- Create a conda env
- conda install numpy ninja pyyaml mkl mkl-include setuptools cmake cffi typing (that installs some requirements)
Then this (which I assume you've already done):
git clone --recursive https://github.com/pytorch/pytorch
cd pytorch
# if you are updating an existing checkout
git submodule sync
git submodule update --init --recursive
And finally set up the conda variable and install:
export CMAKE_PREFIX_PATH=${CONDA_PREFIX:-"$(dirname $(which conda))/../"}
MACOSX_DEPLOYMENT_TARGET=10.9 CC=clang CXX=clang++ python setup.py install
There are issues on the git reporting that behavior here which suggest to add something like:
MACOSX_DEPLOYMENT_TARGET=10.9 CC=clang CXX=clang++ NO_CUDA=1 python setup.py install
Check thatNO_CUDA, this issue as also been mentioned HERE in the forums and it seems to be that it could be an issue caused by the OS and driver versions. If that should be the case I recommend to use Nvidia Docker (hopefully it has mac support) with a pytorch container from https://ngc.nvidia.com/catalog/landing
It that should also fail, your best bet is to install without CUDA support.