Fix dot on MKL builds

[isuruf]

Not sure why MKL_INTERFACE_LAYER used by libmkl_rt.so (linked to by numpy) would affect libmkl_intel_gf.so (linked to by pytorch). It didn't use to. Might be a bug/feature in newer MKL.

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.

I do have some suggestions for making it better though...

For recipe/meta.yaml:

• ℹ️ The recipe is not parsable by parser conda-souschef (grayskull). This parser is not currently used by conda-forge, but may be in the future. We are collecting information to see which recipes are compatible with grayskull.
• ℹ️ The recipe is not parsable by parser conda-recipe-manager. The recipe can only be automatically migrated to the new v1 format if it is parseable by conda-recipe-manager.

_{This message was generated by GitHub Actions workflow run https://github.com/conda-forge/conda-forge-webservices/actions/runs/16190507878. Examine the logs at this URL for more detail.}

[h-vetinari]

Thanks a lot for the fix! I presume this closes #398. I think we should prioritize this over #397, so that people who can't yet use the newer abseil/etc. still get the fix ASAP.

[h-vetinari]

Might be a bug/feature in newer MKL.

Something very strange about this though, because our current pytorch builds have a mkl >=2024.2.2,<2025.0a0 constraint, so it must be a change somewhere else, because mkl 2024 hasn't had any new build recently.

[mgorny]

Thanks, makes sense. The upstream code seems terribly fragile. I mean, apparently it checks whether cblas_sdot() is precise enough, decides that it's not and instead uses fallback that's even less precise?!

[h-vetinari]

Like #396, this is still running into pytorch/pytorch#153737.

Background: In the context of this failure happening in #393, I had found the unmerged pytorch/pytorch#127702, which didn't change anything, and we eventually did conda-forge/triton-feedstock#52; apparently that wasn't a complete fix.

[h-vetinari]

Interestingly, this is still not enough to fix the linalg issues:

+ python -c 'import numpy as np;import torch;x = torch.tensor([2], dtype=torch.complex128);assert torch.dot(x, x).real == 4.0'
Traceback (most recent call last):
  File "<string>", line 1, in <module>
    import numpy as np;import torch;x = torch.tensor([2], dtype=torch.complex128);assert torch.dot(x, x).real == 4.0
                                                                                         ^^^^^^^^^^^^^^^^^^^^^^^^^^^
AssertionError

[Tobias-Fischer]

Should we use np.isclose() instead of == maybe?

[h-vetinari]

Interestingly, this is still not enough to fix the linalg issues:

Brown paperbag moment: when regenerating the patches I unintentionally reverted Isuru's changes 🤦 apologies

[h-vetinari]

Pushed a missing build number bump: 2a435a7

[h-vetinari]

Documenting the windows copy operation (after ensuring old uploads in my channel are not under main label). Filename & ID from here.

$ gh run download 16209749249 --repo conda-forge/pytorch-cpu-feedstock --name conda_artifacts_16209749249_win_64_channel_targetsconda-forge_maincu_hca575dce
$ unzip pytorch-cpu-feedstock_conda_artifacts_.zip
$ cd bld/win-64
$ rm current_repodata.json index.html repodata*
$ ls
libtorch-2.7.1-cuda128_mkl_hca3f899_302.conda       pytorch-2.7.1-cuda128_mkl_py313_h46e6c8c_302.conda
pytorch-2.7.1-cuda128_mkl_py310_h124cda0_302.conda  pytorch-2.7.1-cuda128_mkl_py39_hecd1aea_302.conda
pytorch-2.7.1-cuda128_mkl_py311_h4f3c550_302.conda  pytorch-gpu-2.7.1-cuda128_mkl_h2fd0c33_302.conda
pytorch-2.7.1-cuda128_mkl_py312_hc8193e8_302.conda
$ ls | xargs anaconda upload
$ DELEGATE=h-vetinari
PACKAGE_VERSION=2.7.1
for package in libtorch pytorch pytorch-gpu; do
  anaconda copy --from-label main --to-label main --to-owner conda-forge ${DELEGATE}/${package}/${PACKAGE_VERSION}
done