How to specify hardware specific dependencies

[kunaltyagi]

For a linux system, how do I specify separate wheels for GPU and non-GPU systems? If there's a way to search for executable, I could search for nvidia-smi or rocm-smi in $PATH and then use the cuda/rocm version and the cpu version of the dependency otherwise.

I'm not really sure how that'd look in the pyprojects.toml.

[ofek]

Not possibly yet https://discuss.python.org/t/what-to-do-about-gpus-and-the-built-distributions-that-support-them/7125

[kunaltyagi]

So how do I create my pyprojects.toml such that it has a cpu and gpu version both independent?

I could do the following:

dependencies = [ ... ]

[project.optional-dependencies]
cu118 = ["project_name", "torch=version+cu118"]
cpu = ["project_name", "torch"]

But then even the cpu is optional which doesn't make sense since torch is a hard requirement