❌ Deploy Preview for pytorch-fbgemm-docs failed.

@huydhn has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

Hmm, CPU build fails when this is reverted, so maybe we need this after all This is fixed by using system gcc, it's also gcc-11 now, same as the one from conda.

@huydhn has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

@huydhn has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

I think I get this working for all linux builds now, but the fix is more complicated than I thought. I need to:

1. Use gcc-11 from Nova docker image to fix the GLIBCXX issue.
2. Use CUDA installation from Nova to find all the necessary CUDA libraries there. However, LD_LIBRARY_PATH, CUDNN_INCLUDE_DIR, and CUDNN_LIBRARY env variables need to be updated to point there. It looks like FBGEMM needs to do relocate the wheel similar to what vision does. What do you think? @atalman @q10
3. The CUDA aarch64 build is different than CUDA x86 build because it runs on regular arm runner without NVIDIA gpu, so it doesn't have CUDA installation nor nvidia-smi. This leads to this failure https://github.com/pytorch/FBGEMM/actions/runs/12485012915/job/34843246877#step:18:242. Luckily, FBGEMM already has a logic to fix that, which I reuse. I just realize that there should not be a CUDA aarch64 build at all because with-cuda is set to false there https://github.com/pytorch/FBGEMM/blob/main/.github/workflows/build_wheels_linux_aarch64.yml#L34. This looks like a bug in Nova code, I'm going to push for a fix there.

@huydhn has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

@huydhn has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

Here is the fix to CUDA aarch64 wheel build failure pytorch/test-infra#6112, it shouldn't have been there in the first place

@huydhn merged this pull request in 7d41ee5.