AppleSilicon binaries are build without OpenMP support

[malfet]

🐛 Describe the bug

% python -c "import os;import torch;print(os.uname().sysname, os.uname().machine, torch.__version__, torch.backends.openmp.is_available())"
Darwin arm64 2.1.1 False

Versions

2.1.x, nightly

cc @ezyang @gchanan @zou3519 @kadeng

[garyyaoresearch]

Hi, wondering about what steps you propose to take and how can I contribute to this issue?
Same issue on M2.
print(os.uname().sysname, os.uname().machine, torch.__version__, torch.backends.openmp.is_available()) Darwin arm64 2.2.0a0+gita076a74 False

[malfet]

@garyyaoresearch sorry, edited the title, I've used M1 as alias for Apple Silicon, but yes, they are not the same.
Can you run some CPU benchmarks on your M2 and see whether all cores are used during those tests. If it is indeed the case, perhaps we don't need OpenMP support. (Though OpenMP can still be faster)

[malfet]

Fun fact: it is enabled in nightlies (I assume as result of #110432 ):

% python -c "import os;import torch;print(os.uname().sysname, os.uname().machine, torch.__version__, torch.backends.openmp.is_available())"
Darwin arm64 2.2.0.dev20231129 True

[garyyaoresearch]

@garyyaoresearch sorry, edited the title, I've used M1 as alias for Apple Silicon, but yes, they are not the same. Can you run some CPU benchmarks on your M2 and see whether all cores are used during those tests. If it is indeed the case, perhaps we don't need OpenMP support. (Though OpenMP can still be faster)

Hey no problem. Happy to help :). In running test_sparse.py, nn/test_convolution.py, it seems like all cores are used, but not to full capacity (observed max 350% cpu for 8 cores). In addition, https://discuss.pytorch.org/t/pytorch-unable-to-fully-utilize-cpu-when-performing-svd/192524, indicates such as well.

I think the next steps might be to compare performance against nightly with OpenMP enabled. Do you that will be helpful?

[garyyaoresearch]

Fun fact: it is enabled in nightlies (I assume as result of #110432 ):

% python -c "import os;import torch;print(os.uname().sysname, os.uname().machine, torch.__version__, torch.backends.openmp.is_available())"
Darwin arm64 2.2.0.dev20231129 True

I tried this, but a regular conda install of the nightly doesn't have openmp available. Will try building nightly from binary and also look into the build process.

[huydhn]

So it's not enabled on MacOS nightly for 3.11:

python -c "import os;import torch;print(os.uname().sysname, os.uname().machine, torch.__version__, torch.backends.openmp.is_available())"
Darwin arm64 2.3.0.dev20240205 False

but it's on Linux aarch64:

 python -c "import os;import torch;print(os.uname().sysname, os.uname().machine, torch.__version__, torch.backends.openmp.is_available())"
Linux aarch64 2.2.0 True

[huydhn]

Surprisingly, 3.9 wheel works fine

python -c "import os;import torch;print(os.uname().sysname, os.uname().machine, torch.__version__, torch.backends.openmp.is_available())"
Darwin arm64 2.3.0.dev20240205 True

python -c "import os;import torch;print(os.uname().sysname, os.uname().machine, torch.__version__, torch.backends.openmp.is_available())"
Darwin arm64 2.2.0 True

Need to do a bit of auditing with 3.1[0-2] Only 3.11 binary has this issue. It smells like a bug somewhere in the build here.

[huydhn]

Confirm that this has been fixed on the latest 2.2.1 RC on all python versions