Add `--torch-backend=auto` Flag to `uv add torch`

[NourEldin-Osama]

Summary

Currently, uv pip can automatically install the latest compatible CUDA version of PyTorch via:

1- CLI
uv pip install "torch" --torch-backend=auto --preview

2- pyproject.toml

[tool.uv.pip]
torch-backend = "auto"

It would be great if the same “auto” option were available when adding the torch package directly with uv add.

Proposal:
1- CLI
uv add torch --torch-backend=auto

2- pyproject.toml

[tool.uv]
torch-backend = "auto"

Thanks for this amazing tool!

Example

No response

[FishAlchemist]

This is the PR that initially implemented torch-backend=auto in uv pip install.

• Automatically infer the PyTorch index via --torch-backend=auto #12070

Right now, this is only implemented in the uv pip CLI, since it doesn't quite fit into the lockfile APIs given that it relies on feature detection on the currently-running machine.

So, I think the biggest issue with implementing this feature you're proposing is probably how to represent it in the lockfile, as lockfiles typically don't include dynamic detection content.

[NourEldin-Osama]

@FishAlchemist

Any ideas on how to add this feature and address the challenge of representing dynamic, machine-specific detection (like --torch-backend=auto) in a static lockfile?

Has anyone considered possible approaches or workarounds for this limitation?

I’d like to use uv add with torch-backend=auto so that anyone installing my project gets the appropriate torch version for their CUDA setup.

[peacefulotter]

This would be great!

[intexcor]

You can use wheelnext

[muhammad-fiaz]

this feature throwing me an error, when i add this feature to [tool.uv.pip]

{"torch-backend":"auto","generate-hashes":true,"no-strip-markers":true,"output-file":"requirements.txt","prerelease":"allow","universal":true,"upgrade":true,"verify-hashes":true} is not valid under any of the schemas listed in the 'anyOf' keywordEven Better TOML

this is actually what vscode's warning shows!

[FishAlchemist]

@muhammad-fiaz
I think you can open a separate issue to handle it, as it can easily be overlooked in the comments. As for your question, I have opened an issue (#16545) for you, and I think you can test whether it is working correctly.

[julien-blanchon]

Any update on this ? Could be great to have torch-backend = "auto" with python inline metadata scripts syntax (https://docs.astral.sh/uv/guides/scripts/)

[mon-jai]

@julien-blanchon It is already shipped in v0.9.18.

[noxthot]

@mon-jai
This was not shipped in 0.9.18. As the documentation states:

This setting is only respected by uv pip commands.

This thread is about bringing this functionality to uv add.

I would suggest to reopen the issue, @NourEldin-Osama.

[NourEldin-Osama]

Thanks @noxthot for pointing that out

[kipavy]

and to uv sync