Support nightly PyTorch installs

[charliermarsh]

See: https://github.com/astral-sh/uv/issues/2008.

Commands like pip install -U --prerelease allow torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu121 currently fail, and it's inconvenient to add the +cu121 specifiers where necessary because the user is intentionally pulling in the "latest" release, and so doesn't want to add a dedicated version specifier.

[ewhauser]

Here's a resolution issue I'm running into that pip install handles:

 whisperX @ https://github.com/m-bain/whisperX/archive/befe2b242eb59dcd7a8a122d127614d5c63d36e9.zip
 pyannote-audio @ git+https://github.com/pyannote/pyannote-audio@11b56a137a578db9335efc00298f6ec1932e6317
 torch @ https://download.pytorch.org/whl/cu118/torch-2.2.1%2Bcu118-cp311-cp311-linux_x86_64.whl#sha256=84328a35621cc6a67a182a327baaab67e5f5869981c4b1677ed05f92c15cceb1
 torchaudio @ https://download.pytorch.org/whl/cu118/torchaudio-2.2.1%2Bcu118-cp311-cp311-linux_x86_64.whl#sha256=cd3b1c3582b17792c6d7a367dea0459b123e54d7a4242809ea87ccc10fa220e5
 pytorch_triton @ https://download.pytorch.org/whl/nightly/pytorch_triton-2.1.0%2B7d1a95b046-cp311-cp311-linux_x86_64.whl

> ./uv pip install -r requirements.txt
  × No solution found when resolving dependencies:
  ╰─▶ Because only whisperx==3.1.1 is available and whisperx==3.1.1 depends on torchaudio>=2, we can conclude that all versions of whisperx depend on torchaudio>=2.
      And because only the following versions of torchaudio are available:
          torchaudio<2
          torchaudio==2.2.1+cu118
      and torchaudio==2.2.1+cu118 depends on torch==2.2.1, we can conclude that all versions of whisperx depend on torch==2.2.1.
      And because there is no version of torch==2.2.1 and you require whisperx, we can conclude that the requirements are unsatisfiable.

I believe this is documented here but unclear if there is a way to override?

[charliermarsh]

I think it's possible we could support that specific case by extracting the name and version from the wheel filename (i.e., the URL). Can I ask why you prefer to use direct references for these?

[ewhauser]

This is an existing requirements file so it is possible there is no good reason under uv. However, in order for this to work without direct references, you have to specify:

 --extra-index-url https://download.pytorch.org/whl/cu118
torch==2.2.1+cu118
torchaudio==2.2.1+cu118

which I assume has a performance hit for large requirements files? It does for pip-tools.

I did try that with uv and it started to choke on some of our dependencies that were valid:

 certifi==2024.2.2

 --extra-index-url https://download.pytorch.org/whl/cu118

 whisperX @ https://github.com/m-bain/whisperX/archive/befe2b242eb59dcd7a8a122d127614d5c63d36e9.zip
 pyannote-audio @ git+https://github.com/pyannote/pyannote-audio@11b56a137a578db9335efc00298f6ec1932e6317
 torch==2.2.1+cu118
 torchaudio==2.2.1+cu118
 pytorch_triton @ https://download.pytorch.org/whl/nightly/pytorch_triton-2.1.0%2B7d1a95b046-cp311-cp311-linux_x86_64.whl

  × No solution found when resolving dependencies:
  ╰─▶ Because there is no version of certifi==2024.2.2 and you require certifi==2024.2.2, we can conclude that the requirements are unsatisfiable.

[charliermarsh]

Ah yeah, that's because certifi is published on https://download.pytorch.org/whl/cu118, but only at a different version. We don't allow dependencies to pull "mixed" versions from across indexes as a security measure.

[charliermarsh]

Anyway, I think we can support this specific thing with the URLs.

[charliermarsh]

I created a new issue to track that here: https://github.com/astral-sh/uv/issues/2623

[fernandogrd]

Ah yeah, that's because certifi is published on https://download.pytorch.org/whl/cu118, but only at a different version. We don't allow dependencies to pull "mixed" versions from across indexes as a security measure.

Hi, just experienced this today, specially the different versions of packages being available in the pytorch.

The @ workaround worked, but wondering if there would be other workarounds by now as it is a bit tedious to copy the link of each package specially when there a number of them.

Edit.: Nvm, just found about UV_INDEX_STRATEGY in the docs. I wish pytorch would not keep the copied versions available on pypi though..