cuda 12.0 support request for building pytorch from source code

[yudonglee]

🚀 The feature, motivation and pitch

Motivation:
It does support cuda 12.1, but it does not support cuda 12.0
According to the doc: https://github.com/pytorch/pytorch#from-source , the magma-cuda* that does not match the CUDA version 12.0 from https://anaconda.org/pytorch/repo. Only magma-cuda121 can be found, no magma-cuda120

feature: please support cuda 12.0 for building pytorch from source code

Thank you!

Alternatives

No response

Additional context

No response

cc @malfet @seemethere @ngimel

[malfet]

@yudonglee please do not hesitate to propose a PR against https://github.com/pytorch/builder to enable CUDA-12.0 magma builds.

But just curious, why CUDA-12.0 builds will be beneficial when 12.1 is already available?

[ExtReMLapin]

@yudonglee please do not hesitate to propose a PR against https://github.com/pytorch/builder to enable CUDA-12.0 magma builds.

But just curious, why CUDA-12.0 builds will be beneficial when 12.1 is already available?

There are some systems with very specific drivers versions, for example nVidia GRID drivers, where are not always available in all versions on all systems.

For example here we're stuck on drivers with cuda support for 12.0 only.

[yudonglee]

@yudonglee please do not hesitate to propose a PR against https://github.com/pytorch/builder to enable CUDA-12.0 magma builds.

But just curious, why CUDA-12.0 builds will be beneficial when 12.1 is already available?

On my computer environment, it's not available to update cuda's version to 12.1

Anyway, I still have any clue on how to fix this. Need some help here :)

[bryantbiggs]

xref pytorch/builder#1759

but if I am reading the builder repo correctly, the version of magma used is quite dated and does not support CUDA12 or Hopper arch. I don't know what "support" means in this context, otherwise if it truly wasn't supported I imagine an issue would have been raised earlier for this