Arch Build Issues: provide pypi installs?

[nvidia]

Arch Build Issues: provide pypi installs?

[mcarilli]

Weird, some of the errors look like you're attempting to compile c++14 features with a c++11 compiler (e.g. error: body of constexpr function...). Apex (and Pytorch) should only require c++11. It also looks like the build is pulling in headers for gcc 6.4.1. I don't think this is an issue specific to Apex. Did you build Pytorch from source? Also, are you able to compile their Pytorch c++ extension tests? Try

cd <pytorch_repo_root>/test/cpp_extensions/
python setup.py install

Moving forward, we are working on a pip-installable version of Apex, as well as a Python-only installation. The C++ backend is not required by FP16_Optimizer and apex.parallel.DistributedDataParallel, and Amp only requires the backend as a performance optimization.

[mcarilli]

Apex is now a Python-only package for simplicity and robustness. We're working on moving Apex's former C++ backend into Pytorch itself. In the meantime, Amp, FP16_Optimizer, and apex.parallel.DistributedDataParallel continue to work, although Amp may exhibit mild (~2%) performance reduction. This performance hit is temporary and will be reversed once the C++ backend is available for Amp to use through Pytorch core.

If you have Pytorch on your system, you should be able to install and use Apex without any more C++ build problems.

[jramapuram]

Great, thanks very much! Will give this a shot and report back. Looking forward to Amp integration in pytorch: will make fp16 training a whole lot easier. Any idea on the timeline for this? I can pull pytorch and compile it if the PR has been accepted into master.

[jramapuram]

This works perfectly now. pip install works; the build works as well. I had some issues with an AUR version of gcc-6 on arch causing the above conflicts. User error :)