Migrate build data to pyproject.toml

[matthewdouglas]

• Moves most metadata to pyproject.toml, conforming with PEP 517, PEP 518, and PEP 621
• Removes requirements.txt files (but not the conda environment.yml ones just yet)
• Updates docs to instruct on usage of editable install for development (see: PEP 660)
• Updates workflow to skip unnecessary dependency installation for building wheels

[github-actions]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[rickardp]

Don't we need any version constraint at all? Maybe at least a minimum version?

[matthewdouglas]

I do think we should have at least minimums, but I'm not sure what they should be. I took these runtime dependencies out of setup.py just as they were.

If it were up to me, I would pin torch>=1.13.0 on the low end for all platforms/backends assuming everything works. Why?

• Oldest release which did not ship any wheels with CUDA < 11
  • Migrated to CUDA 11.6 and 11.7, which is right in line where I think bitsandbytes cutoff should be [1]
  • Apple silicon support (MPS) promoted from experimental/prototype to beta
    • 🤗 Accelerate docs "strongly recommend" torch>=1.13 for MPS
• AWS has HuggingFace containerswith the oldest having torch==1.13.1+cu117 on py3.9
  ** In general, all of the Python 3.9+ containers with PyTorch appear to be torch>=1.13.0
• Azure Container for PyTorch still has a 1.13. configuration

But it could also be a little more permissive for some environments; e.g. transformers recently adopted a constraint of torch>=1.11,!=1.12.0 in huggingface/transformers#28207, which was also adopted in UKPLab/sentence-transformers#2432.

[rickardp]

IMHO (and I recognize this might ba a bit controversial) we should set minimum to what we test on, then pin to these when we test. I don't think we should overstretch to keep supporting old versions of dependencies at the expense of stability or complexity. So I would say ideally we set minimum to what we currently build on and either no upper bound or < next major.

As a user of any package I prefer to know what works or not and not having to be the guinea pig.

But this is my personal opinion, not sure what the rest of you think.

[matthewdouglas]

Ideally we would have a CI setup that could test a wider matrix so we can be confident about how we pin these, for sure.

For me part of the point of this bitsandbytes project is to try and make things accessible for a wide range of users - so that's across hardware, software, and user's expertise. Since many people, particularly in academia, don't have much control over things like OS, python version, or even torch version, we've got to support a large range in order to make things easy for users. That does come at the cost of maintenance, obviously. Need to find a good balance here.

So at least with this change it's being restricted to torch>=1.11,!=1.12.0, which is in line with transformers. I don't know that all of those versions of torch will work, but it's an incremental step forward compared to install_requires=["torch", "numpy"], and given the lack of issues around torch versions it does seem likely that it would work.

[matthewdouglas]

Rebased

[Titus-von-Koeller]

Rebased

This is really weird, somehow the diff of the PR got all messed up now. This makes it really hard to review.

It seems to me this PR is otherwise ready for review, right?

[matthewdouglas]

Rebased

This is really weird, somehow the diff of the PR got all messed up now. This makes it really hard to review.

It seems to me this PR is otherwise ready for review, right?

Not sure what happened there but I think it should look more reasonable now. Apart from that, yes, it's ready.