Gentoo portage overlay with backported ebuilds to support CUDA 10.2
This overlay contains Gentoo packages taken from the main tree - sometimes with backports - to allow running CUDA 10.2 with an otherwise up to date package database.
I'm fully aware that his is a Don Quixote effort as more and more patches will be required as time progresses. NVIDIA, for example, have just recently dropped support for their 470 series Linux kernel driver. But for now it looks as if it's still possible to number crunch ML models with fairly ancient GPU and a fairly modern operating system.
My particular aim is to support my NVIDIA GeForce 750M card with as little ancient packages as possible. When backporting I have tried to rely on existing commits in pytorch and related packages whenever possible. Originally I had ported ONNX 1.8.1 but then I found a single pytorch commit that would allow me to compile it against a recent version. I keep the old port in the overlay in case anyone has a specific concern and wants to base their work on my backport.
pytorch 1.7.0 requires python 3.10. The commits in the pytorch repository that lift pytorch to be compatible with python 3.11 are very complex. Depending on python 3.10, which is said to be significantly slower than more recent versions, was the most viable path forward.