pip vs. pixi inconsistencies with a corporate artifactory

[geoHeil]

I am struggling with a corporate artifacts mirror

The following (plain pip) interaction with the artifactory mirror works fine

pip uninstall -y torch
pip install torch --index-url https://bin.mycorp.com/artifactory/api/pypi/pytorch-cu129-pypi-remote/simple

When instead setting this from pixi

Error:   × failed to solve the pypi requirements of environment 'myenv' for platform 'linux-64'

  ├─▶ failed to resolve pypi dependencies

  ╰─▶ Because torch was not found in the package registry and you require torch>=2.8.0,<3, we can conclude that your requirements are unsatisfiable.

Installing the plain pip torch default package from pypi works - but does not allow to select the desired CUDA version. Strangely

The pixi setup

[tool.pixi.feature.torch -cpu.pypi-dependencies]
# mirror of https://download.pytorch.org/whl/cpu

torch = { version = ">=2.8.0,<3", index = "https://bin.mycorp.com/artifactory/api/pypi/pytorch-cpu-pypi-remote/simple" }


[tool.pixi.feature.ai-example-cuda]
platforms = ["linux-64"]
system-requirements = { cuda = "12.9" }

 
[tool.pixi.feature.torch -cuda.target.linux-64.pypi-dependencies]
# mirror of https://download.pytorch.org/whl/cu129

torch = { version = ">=2.8.0,<3", index = "https://bin.mycorp.com/artifactory/api/pypi/pytorch-cu129-pypi-remote" }

Notice: this is only a problem for the https://bin.mycorp.com/artifactory/api/pypi/pytorch-cu129-pypi-remote mirror the CPU one https://bin.mycorp.com/artifactory/api/pypi/pytorch-cpu-pypi-remote/simple. However, pip works just fine with the CUDA mirror

[tdejager]

Hmm can't be sure, what included in menv and does -v show anything of note, does it query the correct index?

[geoHeil]

basically solo this part that you see on the screen.

In fact if I change (delete the index and use the pypi default publiication) things work/resolve fine

[geoHeil]

no I could not yet spot anything useful in the verbose resolution procedure