Apple support/build instructions

[YannickLimmer]

Hi, I figured out how to run the latest version of this package on MacOS. I thought I share this with you.

[patrick-kidger]

Great, thanks!

I think it should be possible to still require OpenMP, though. (We used to have Mac builds with OpenMP.) Either way I don't want to remove the ifndef statement. (And instead require users to modify the source code, as you're currently instructing.)

Could you add a disclaimer at the top that these are experimental/not-officially-supported instructions?

Also I don't think the business with site should be necessary - you should be able to clone anywhere.

[hans-riess]

Thanks for the instructions @patrick-kidger. I'm personally having some trouble still, but it could easily be just me!
torch=2.2.0.dev20231018
python=3.11.5
I'm running macOS Sonoma 14.2 on M1 Pro. If you have any further tips, greatly appreciated!

[jesuinovieira]

I ran into similar issues when building the PyTorch-2.x update branch on macOS (M3). I first tried the approach here (using LLVM and libomp); the package built successfully but failed at import, likely because it was compiled with LLVM’s libc++ while Python/torch use Apple’s toolchain.

After removing all OpenMP-related flags (-fopenmp, -lomp, -Xpreprocessor) and rebuilding with the default Apple clang, the extension compiled and imported correctly.