latest AMI `ami-0285bfb786db80af3` in Actions does not have `nvidia-smi`

[ilan-gold]

So I am really new to this, please forgive any mistakes.

I checked the https://github.com/scverse/gpu-ci-config/actions/workflows/build-ami.yml section and found the latest image as ami-0285bfb786db80af3 which I then plugged into scverse/anndata#2269 but nvidia-smi didn't work.

I also tried running the "Update AMI" CI job as per the README.md and that errored out with a 403 on making the PR: https://github.com/scverse/gpu-ci-config/actions/runs/20598977237/job/59159801996 but I don't expect there to be any difference in the outcome (broken nvidia-smi).

As a final check, I tested the image using https://github.com/scverse/gpu-ci-config/actions/runs/20599052838/job/59160046962

I must be missing some step of the process. What's up here?

[grst]

The failing PR is related to missing token permissions that I screwed up when renewing the token. I'll look into that.

I can't help with why nvidia-smi doesn't work though. @Intron7 @aktech

[ilan-gold]

Just to be clear, this blocks supporting cupy-cuda13x in CI. @Intron7 Could you have a look at this? Anyone getting a new version of torch will have a broken cupy installation with cupy-cuda12x: https://github.com/pytorch/pytorch/releases/tag/v2.11.0

It would be awesome to be able to test out cupy-cuda13x to stay up with torch

[flying-sheep]

ping @Intron7