There are a few obvious roadblocks:

• There are a huge number of bundled dependencies in third_party/, which would have to be unbundled or carefully handled in accordance with guidelines
• Most people would want to use it with NVIDIA GPUs, but the CUDA Toolkit is not free software, so support would have to be entirely disabled
• Intel MKL is also not free software and would have to be disabled
• AMD GPU support via ROCm is theoretically possible, but that stack is currently only partially packaged in Fedora
• The codebase is very large and fast-moving; keeping up would not be easy

Basically, it might be possible, but it would be a huge initial packaging and maintenance undertaking, and the result would be a CPU-only build, or perhaps someday a CPU-and-AMD build.

Thanks for the summary @music. I opened this ticket primarily to get the info all in one place. @ankursinha's comment in #254 already made me think someone had looked into it in the past.

If nothing more, we could close wontfix or keep it open as on hold and use it as a sink for other software that has direct or indirect dependencies on PyTorch. Then at least it's documented why certain software will not be available in Fedora.

Yes, I think it’s good to have an issue (open or closed) for discussion and technical details, since pytorch comes up a lot.

There's this page that documents software that we cannot include in Fedora and why:

https://docs.fedoraproject.org/en-US/neurofedora/excludedtools/

The nvidia drivers are already there, so folks should sort of know that anything that requires cuda etc cannot be included, but we could list pytorch etc. explicitly if that'll make it clearer.

Commit c11fa20a fixes this issue

Pagure gets links wrong. Here's the commit:

https://pagure.io/neuro-sig/documentation/c/c11fa20ac1f70e77e9636dcbe3f6380fb67190eb?branch=main