GPU / Nvidia option for Production Branch rather than Feature Branch

[JohnAtl]

One issue I've had across immutable distros (UBlue, and others, the latest being Vanilla OS) is that the installations include Nvidia's New Feature Branch (currently 530.41.03). If one wants to do machine learning-/AI-type work, the cuda, cudnn, etc. libraries are still on the Production Branch (currently 525.116.04). This leads to a lot of hoop-jumping to get the machine learning/AI stack to work.
If one wishes to try to compile solutions to this dilemma in a Distrobox, another set of complications arise in that the Distrobox OS likely has a different kernel version as the host OS, thus the header files do not match.
While this may be a "me" problem, I feel like there are other people with similar workflows involving Nvidia and perhaps other drivers with similar challenges, and wanted to post this for discussion.

[mister-drgn]

This is something I've been meaning to try, but I haven't yet installed ublue on a work computer. Are you saying you would get an error if you use distrobox to set up a container from an nvidia cuda image (https://hub.docker.com/r/nvidia/cuda)? This works fine for me on a regular Fedora install, but I'm not sure whether I'm using the 525 or 530 driver.

[castrojo]

If I recall the reason we couldn't keep 525 is that rpmfusion dropped it? It'd be ideal if we could keep everything. 😶‍🌫️

[mister-drgn]

Various websites (e.g., https://www.cloudbooklet.com/how-to-set-up-deep-learning-architecture-on-ubuntu-22-04/) suggest to me that the 530 drivers are compatible with cuda and cudnn, but perhaps only with the latest version, cuda 12.1. This could be an issue for ML because much of what's out there has been developed with older cuda versions.

But isn't there also an option to install the 470 driver?