Compiling torch for MacOS 13 Ventura (M1)

[cdmatters]

I'm trying to compile pytorch v.1.13.0 from source for MacOS (yes CPU only, which simplifies things) and we seem to get reaaally close before you hit something that looks pretty critical.

Is there any chance you happen to have a branch of any modifications you made to pytorch to make it compile for your custom package index? Or does this require something a little more involved?

[cdmatters]

Ah okay I have found your personal fork of pytorch with the nogil branch. I'll try here!

[colesbury]

Yeah, that’s the right branch. I haven’t built 1.13 or macOS wheels yet. I’d like to do so, but I won’t have time to look into it until next week.

[cdmatters]

I'm actually happy with 1.9.0 but I seem to be setting segfaults building sleef

[cdmatters]

This is the error I'm encountering, dubiously no comment since 2021. The comment suggests cross compile should be fine, so maybe wheels will be alright...

[colesbury]

I'm looking into this now. I might end up building 1.13 wheels because I'm not sure 1.9.0 supports M1.

[colesbury]

I'm still working on building the wheels, but if you are feeling impatient you can try:

git clone https://github.com/colesbury/pytorch.git -b v1.13.1-nogil

And once you install the prerequisites, build it with:

BUILD_FUNCTORCH=0 python setup.py install

Functorch doesn't build right now because of changes to the bytecode format and PyCodeObject structure.

[colesbury]

I've uploaded PyTorch 1.13.1 wheels for macOS (x86-64 and arm64). Please let me know if you run into any issues.