`transformers-cli` now requires PyTorch

[setu4993]

System Info

N/A

Who can help?

@gante (author of #35714)

Reproduction

uv run --python 3.12 --with transformers transformers-cli --help

Expected behavior

Since #35714 (released in v4.49.0), the transformers-cli now requires PyTorch to even run a simple: transformers-cli --help command.

I am pretty sure it is originating from this line.

My question is if this is intentional or unintentional. So far, PyTorch has been an optional dependency for any functionality with CLI or the package itself.

[Rocketknight1]

Seems real, and I don't think we should require Torch for transformers-cli. I think this is probably a testing failure more than anything, though - cc @gante, maybe we should have tests for transformers-cli that don't @require_torch so we catch things like this in future?

[scw]

I agree this is an issue that should be addressed — two dependencies that aren't installed without adding some extras elements, pytorch and rich, are now required to run transformers-cli. Helpfully, we get a message about PyTorch not being found and how that affects functionality, but the chat.py component isn't wrapped in a try / except to prevent import failure with the default requires list.

[machinelearningprodigy]

I think you're running into #35714, which made PyTorch a required dependency for transformers-cli.
Even a simple transformers-cli --help now needs PyTorch to work.
You can fix this by installing PyTorch (pip install torch) or downgrading transformers to a version before v4.49.0.

[Rocketknight1]

gentle ping @gante! Was that dependency intentional?

[gante]

@Rocketknight1 not intentional, I'll add a guard (and tag the branch for a patch release)