Put back MPS builds

[NicolasHug]

(this is a follow-up and more up-to-date version of #8456)

The M1 CI jobs were broken for ~1 week (#8456) and it turns out the problem was caused by the MPS build. We deactivated the MPS builds in #8472 and the M1 jobs (all using macos-m1-stable) are now green.

We have to put back the MPS build before the release though, otherwise torchvision won't provide MPS-compatible custom ops.

In #8476 (macos-m1-stable), #8473 (macos-m1-13) and #8477 (macos-m1-14) I'm trying to add back those MPS builds, but they all fail with the same error as previously seen back in #8456:

  File "/Users/ec2-user/runner/_work/vision/vision/pytorch/vision/test/smoke_test.py", line 7, in <module>
    import torchvision
  File "/Users/ec2-user/runner/_work/vision/vision/pytorch/vision/torchvision/__init__.py", line 10, in <module>
    from torchvision import _meta_registrations, datasets, io, models, ops, transforms, utils  # usort:skip
  File "/Users/ec2-user/runner/_work/vision/vision/pytorch/vision/torchvision/_meta_registrations.py", line 164, in <module>
    def meta_nms(dets, scores, iou_threshold):
  File "/opt/homebrew/Caskroom/miniconda/base/envs/ci/lib/python3.10/site-packages/torch/library.py", line 653, in register
    use_lib._register_fake(op_name, func, _stacklevel=stacklevel + 1)
  File "/opt/homebrew/Caskroom/miniconda/base/envs/ci/lib/python3.10/site-packages/torch/library.py", line 153, in _register_fake
    handle = entry.abstract_impl.register(func_to_register, source)
  File "/opt/homebrew/Caskroom/miniconda/base/envs/ci/lib/python3.10/site-packages/torch/_library/abstract_impl.py", line 30, in register
    if torch._C._dispatch_has_kernel_for_dispatch_key(self.qualname, "Meta"):
RuntimeError: operator torchvision::nms does not exist

CC @malfet @huydhn

[NicolasHug]

From #8456 (comment)

Though it fails on my Mac even with MPS disabled, so I have no idea what is going on...

@malfet are the tests failing on your Mac with the same error message on main? 🤔

[huydhn]

I could reproduce the issue locally and can confirm that disable MPS makes import torchvision work again. However, I think the recent upgrade to MacOS 14 might be a red herring because using macos-m1-13 still fails in the same way, i.e. https://github.com/pytorch/vision/actions/runs/9425764020/job/25967868583?pr=8485

[huydhn]

This issue seems to appear in PyTorch commit between https://hud.pytorch.org/hud/pytorch/pytorch/d66f12674cfe0151a86dc10b8de216f83bf42e6e (failed) and https://hud.pytorch.org/hud/pytorch/pytorch/0ff2f8b52248323bbe25108b64e706c43390cb72 (success). I need to narrow this range down a bit.

[huydhn]

I think I have found the change responsible for this in pytorch/pytorch#127265. I have reverted it so that we can test it out in the next PyTorch nightly build. Reverting it locally fix the issue for me.

[huydhn]

Testing in #8485, but pending the revert https://hud.pytorch.org/hud/pytorch/pytorch/75b0720a97ac5d82e8a7a1a6ae7c5f7a87d7183d to get into the next PyTorch nightly build.

[NicolasHug]

#8485 worked, thanks a ton @huydhn !