adopt https://astral.sh/blog/wheel-variants for distribution

[sergei-something-ai]

Hello!

I was looking today at the sglang distribution model.
I started at pypi with sgl-kernels and trying to piece together what should I do to get some fresh version of it

https://pypi.org/project/sgl-kernel/#files

Then I read a number of issues that raise various problems with packaging and edge-cases and lack of wheels for arm and so on. (i.e. #7504 (comment))
I also looked at the GH action workflow for wheels publishing a little bit trying to wrap my head around it.

The problem is very complex indeed and python doesn't have a great way to deal with such accelerator-specific wheels. Yet!

But there is a promising project
https://wheelnext.dev/proposals/pepxxx_wheel_variant_support/

and more importantly you can use experimental build of uv today to run uv pip install torch and it will do "THE RIGHT THING for your accelerator"

Here are more details
https://astral.sh/blog/wheel-variants

TL;DR: In collaboration with PyTorch, NVIDIA, and Quansight, we're releasing an experimental build of uv with support for "wheel variants". Variants make it easier to distribute and install Python packages built for diverse hardware or software setups, like GPUs or SIMD support.

There is a lot of momentum behind this design, even though it's not yet accepted as a PEP.
I though that it could be a good idea to start looking into adopting WheelNext format.