Fix pytorch-cpu/gpu build only for single python version

[jeongseok-meta]

Resolves #281

Checklist

• Used a personal fork of the feedstock to propose changes
• Bumped the build number (if the version is unchanged)
• Reset the build number to 0 (if the version changed)
• Re-rendered with the latest conda-smithy (Use the phrase @conda-forge-admin, please rerender in a comment in this PR for automated rerendering)
• Ensured the license file is being packaged.

[conda-forge-admin]

Hi! This is the friendly automated conda-forge-linting service.

I just wanted to let you know that I linted all conda-recipes in your PR (recipe/meta.yaml) and found it was in an excellent condition.

[jslee02]

aarch64 + cuda 12.0 build failed: linux_aarch64_c_compiler_version12c_stdlib_version2.28cuda_compilercuda-nvcccuda_compiler_version12.0cxx_compiler_version12-log.zip

Could not solve for environment specs
The following packages are incompatible
├─ �[31mpytorch-gpu 2.5.1.* cuda120hf733e3a_202�[0m is not installable because it requires
│  └─ �[31mpytorch 2.5.1 cuda_generic*202�[0m, which does not exist (perhaps a missing channel);
└─ �[31mpytorch 2.5.1.* cuda_generic*202�[0m does not exist (perhaps a typo or a missing channel).

[hmaarrfk]

here goes an other 2 hours.....

[jslee02]

Still building, but it seems to work! Spoken too early. It's pytorch:

and pytorch-cpu still outputs for only one python version:

Build log: log_linux64_mkl_cpu.zip

[jeongseok-meta]

Wait, perhaps this is the expected behavior (not specific to any python version). If so, it appears to be working as intended. I'll continue building...

[jeongseok-meta]

Still building, but the estimated time to completion is 4-5 hours.

Packages that have been built so far:

• https://anaconda.org/jslee02/libtorch
• https://anaconda.org/jslee02/pytorch
• https://anaconda.org/jslee02/pytorch-cpu
• https://anaconda.org/jslee02/pytorch-gpu

[hmaarrfk]

Wait, perhaps this is the expected behavior (not specific to any python version). If so, it appears to be working as intended. I'll continue building...

Correct, this is the expected behavior.

When the megabuild happens a single pytorch-cpu will be made that can point to any pytorch for any version of python.

The old used exact which doesn't work well for "splayed" builds where we have:

1. C++ build.
2. Many python builds

[hmaarrfk]

I'm sorry to say this, but I think it might be good to include this patch:
pytorch/pytorch@a00faf4

It is manifesting as a failure in conda-forge/torchvision-feedstock#102

[danpetry]

I think you might want to remove this line, otherwise you'll fix to a python version for non-megabuild variants?

[hmaarrfk]

The non megabuillds should have multiple packags created. Is that not the case?

[danpetry]

ah, ok. I'm sure you're right, I need to check more carefully I think.

[danpetry]

Still getting familiar with your recipe.

[hmaarrfk]

The megabuilds were created since before the large runner from quantsigh hour builds, we used to HAVE to build locally. And it would effectively save me a 12 hours of rebuilding libtorch 4 times.

[danpetry]

Thanks. I guess "megabuild" = 1 x libtorch and 5 x python bindings/code, as opposed to 5 x normal builds (libtorch + bindings + python code)?

[jslee02]

The build is done!

Log: log_files_1.zip, log_files_2.zip
Pkgs (only new files):

• https://anaconda.org/jslee02/libtorch
• https://anaconda.org/jslee02/pytorch
• https://anaconda.org/jslee02/pytorch-cpu
• https://anaconda.org/jslee02/pytorch-gpu

[hmaarrfk]

anaconda copy --from-label main --to-label main --to-owner conda-forge jslee02/libtorch/2.5.1
anaconda copy --from-label main --to-label main --to-owner conda-forge jslee02/pytorch/2.5.1
anaconda copy --from-label main --to-label main --to-owner conda-forge jslee02/pytorch-gpu/2.5.1
anaconda copy --from-label main --to-label main --to-owner conda-forge jslee02/pytorch-cpu/2.5.1