Simplify Nvidia dependencies

[andyl]

For a given version of EXLA, I find it difficult to know which combination of CUDA drivers, CUDA runtime and cuDNN are needed to make everything work.

Just posting an idea here: Could EXLA/XLA install these dependencies automatically? Or emit a bash script to set things up properly? Or include some sort of healthcheck task?

[polvalente]

Other than looking in the README for the https://github.com/elixir-nx/xla?tab=readme-ov-file#xla_target table, which does specify the expected versions, I believe automatically installing these would be too invasive.

As a personal anectode, every time my Linux kernel updates, I have to tweak the NVIDIA driver installation -- and the easiest way is to reinstall the NVIDIA drivers and CUDA. If EXLA started tampering with that as well, it could become an unstable system.

I think we could, however, provide some sort of mix xla.check_version --target cuda12 or something similar, where we validate the dependencies for each of the given lines, and emit a warning/fail the task when the validation fails.

[andyl]

Thanks @polvalente a check_version task would be great! Would you like me to submit a PR for it? If so, LMK if there are features you would like, or guidelines you would like me to follow.

[polvalente]

I'd like @jonatanklosko to chip in here as he might have some extra thoughts, before we move on with a PR.

I think the task should accept the same values as XLA_TARGET, or use its value directly, and then based on that, validate the key dependency versions.

However, I'm not sure if this is actually doable in a portable way for all of the supported targets.

[jonatanklosko]

I think portable checks would be problematic. For example, to find out cuDNN version, I'm not sure if there's a more reliable way than searching installed system packages, which is OS-specific. Perhaps we could look for certain files, but I'm pretty sure that's going to be too brittle.

Maybe we should make the EXLA->XLA navigation version aware. Currently the EXLA docs link to https://github.com/elixir-nx/xla, which is always main. We could instead link to `XLA` and then include the versions information in the XLA module docs. Or perhaps that's still too much and we should move all information to the EXLA docs altogether. @josevalim @polvalente thoughts?

[polvalente]

I like the idea of linking to XLA and having the version-aware info there!

[jonatanklosko]

I opened PRs with docs improvements, I'm going to close this.