Improving reproducibility: Moving to UV

[KennethEnevoldsen]

We are already using ruff (instead of flake8 and/or black) it might similarly be great to move to uv. Uv is made by the same (great) team.

A benefit would be that we can add a .lock-file to allow for more reproducibility. This will also make some of the CI easier going forward.

uv is pip compatible so it would still be possible to use pip without issues (as far as I am aware).

To start the discussion I will just add a few people:
@isaac-chung, @orionw, @Muennighoff, @x-tabdeveloping

[KennethEnevoldsen]

(@MartinBernstorff might also have a few comments to add)

[Samoed]

uv is a good tool, but it doesn't support development dependency groups, which is a minor issue in my opinion. I use pdm, which has a similar CLI to uv and can use uv as a backend, though this is experimental.

[KennethEnevoldsen]

are you referring to stuff like pip install mteb[dev]?

[mrshu]

@KennethEnevoldsen it is a similar concept, although what you refer to is called "optional dependencies", as far as I know.

This tracking GH issue has a bit more detail: astral-sh/uv#5632

I would indeed call it a very minor issue.

[Samoed]

Yes, @mrshu right. PDM info

[KennethEnevoldsen]

Thanks for clarifying seems like we can at least keep the current setup for now (without loosing anything) and potentially switch to dependency groups going forward

[orionw]

As someone who has never used ‘uv’ I would just suggest that we have very good documentation for it if we do switch.

My biggest hesitation is making it harder for new collaborators to have to learn a new package manager. I’ve heard great things about it, but TBF I had heard that about poetry and it was fairly non-intuitive to use, whereas everyone knows how to use pip. But I haven’t used uv so perhaps it actually is easy :)

Edit: I see that it is pip compatible so maybe not an issue.

[KennethEnevoldsen]

Yea while poetry requires a complete redo of the pyproject.toml, uv simply adds a uv.lock and otherwise I don't believe we need to see a lot of changes.

One benefit of the uv.lock is that people would be able to see which exact versions people have got things to run with at a specific time.