migrate from `setup.py` to `pyproject.toml`

[jorenham]

I'd also recommend uv for pyproject.toml-based project management. That would also allow getting rid of requirements-dev.txt, simplifying the project setup.

_{But something tells me @lucascolley is going to bring up pixie as an alternative to uv.}

[lucascolley]

uv would be good! Buuut... since we are dealing with GPU packages, we are going to want to use libraries like CuPy from conda-forge rather than PyPI (at least for CI, Ralf brought this up at #251 (comment)). At that point, it's probably worth going all in on Pixi.

If somebody wants to contribute a uv setup, I think that would still be a useful contribution! But I don't really have a horse in this race as I rarely develop array-api-compat locally. We can always change stuff like this in the future anyway.

[jorenham]

Pytorch can be used with uv, and they even have a guide for it: https://docs.astral.sh/uv/guides/integration/pytorch/#installing-pytorch

Jax also shouldn't be a problem, and can nowadays even be used without https://github.com/jorenham/jax_pep503

If I remember correctly, I managed to get dask working within a uv project as well.

And I know for certain that numba works as well.

---

That's all of the gpu-related stuff, right?

[jorenham]

If somebody wants to contribute a uv setup, I think that would still be a useful contribution! But I don't really have a horse in this race as I rarely develop array-api-compat locally. We can always change stuff like this in the future anyway.

I'll give it a try

[lucascolley]

For sure, you can install and use GPU packages from PyPI. But there are various reasons why, especially from a development perspective, it is better to work with a package manager/distribution which is aware of native dependencies. There is more on this at https://pypackaging-native.github.io/key-issues/gpus/.

[lucascolley]

Also see #255 before starting work on this @jorenham

[jorenham]

For sure, you can install and use GPU packages from PyPI. But there are various reasons why, especially from a development perspective, it is better to work with a package manager/distribution which is aware of native dependencies. There is more on this at pypackaging-native.github.io/key-issues/gpus.

Yea that makes sense. But in this project, does that really matter?

[lucascolley]

I suppose it probably doesn't make a big difference, but just better to follow best practices given the choice 🤷‍♂️

[NeilGirdhar]

Can this be closed by #255?

[ev-br]

Indeed, thanks @NeilGirdhar !