uv pip install a wheel with extra index url doesn't resolve, normal pip works in fresh env

[rbavery]

• A minimal code snippet that reproduces the bug.

I've attached a .zip containing the cpu wheel I'm trying to install with uv.

this works find in a fresh conda env with python 3.10

pip install sedonaai_cpu-0.1.1-py3-none-any.whl --extra-index-url https://download.pytorch.org/whl/cpu

But uv fails to resolve in a fresh venv with python 3.10

→ uv pip install ./sedonaai_cpu-0.1.1-py3-none-any.whl --extra-index-url https://download.pytorch.org/whl/cpu --prerelease=allow
  × No solution found when resolving dependencies:
  ╰─▶ Because torch==2.2.0 is unusable because no wheels are available with a matching Python implementation and torch==2.2.1 is unusable because no wheels are available with a matching
      Python implementation, we can conclude that any of:
          torch==2.2.0
          torch==2.2.1
       cannot be used.
      And because torch==2.2.2 is unusable because no wheels are available with a matching Python implementation, we can conclude that any of:
          torch==2.2.0
          torch==2.2.1
          torch==2.2.2
       cannot be used. (1)

      Because only the following versions of torchvision are available:
          torchvision<=0.17.0
          torchvision==0.17.0+cpu
          torchvision==0.17.1
          torchvision==0.17.1+cpu
          torchvision==0.17.2
          torchvision==0.17.2+cpu
          torchvision>=0.18.0
      and torchvision==0.17.0 is unusable because no wheels are available with a matching Python implementation, we can conclude that torchvision>=0.17.0,<0.17.0+cpu cannot be used.
      And because torchvision==0.17.0+cpu depends on torch==2.2.0, we can conclude that torchvision>=0.17.0,<0.17.1 depends on torch==2.2.0.
      And because torchvision==0.17.1 is unusable because no wheels are available with a matching Python implementation and torchvision==0.17.1+cpu depends on torch==2.2.1, we can
      conclude that torchvision>=0.17.0,<0.17.2 depends on one of:
          torch==2.2.0
          torch==2.2.1

      And because torchvision==0.17.2 is unusable because no wheels are available with a matching Python implementation and torchvision==0.17.2+cpu depends on torch==2.2.2, we can
      conclude that torchvision>=0.17.0 depends on one of:
          torch==2.2.0
          torch==2.2.1
          torch==2.2.2

      And because we know from (1) that any of:
          torch==2.2.0
          torch==2.2.1
          torch==2.2.2
       cannot be used, we can conclude that torchvision>=0.17.0 cannot be used.
      And because sedonaai-cpu==0.1.1 depends on torchvision>=0.17.0, we can conclude that sedonaai-cpu==0.1.1 cannot be used.
      And because only sedonaai-cpu==0.1.1 is available and you require sedonaai-cpu, we can conclude that the requirements are unsatisfiable.

• The command you invoked (e.g., uv pip sync requirements.txt), ideally including the --verbose flag.

the logs wth verbose are attached
logs.txt

artifact.zip

[zanieb]

Could you check if this is related to a known incompatibility with pip?

There are a lot of discussions about behaviors with extra index URLs:

• wheel on pypi.org not found when using --extra-index-url that includes non-compatible wheels #2542
• Question: How to handle multiple private extra indexes? #2310
• Add support for pinning a package to a specific index #171
• uv fails to use extra index url #1377
• uv pip install for package works from --index-url but not from --extra-index-url #1451
• --extra-index-url failing when package not in --index-url (polygraphy 0.47.1) #1600

[charliermarsh]

My guess is that this is related to our limitations around local version identifiers: https://github.com/astral-sh/uv/blob/main/PIP_COMPATIBILITY.md#local-version-identifiers.

Can you look at the versions of Torch and Torchvision that pip is installing, and add them as direct dependencies? Like:

pip install sedonaai_cpu-0.1.1-py3-none-any.whl torchvision==0.17.0+cpu torch==2.0.0+cpu --extra-index-url https://download.pytorch.org/whl/cpu

[charliermarsh]

Closing for now in the absence of more info, but let me know if you have other questions.