AMD GPU mount /dev/dri in workload container

[bernardgut]

Bug Report

This is actually a continuation of this : #4001 but since that thread is locked I am starting a new one.

I have a bunch of workloads similar to plex on bare metal nodes with AMD GPUs. I see a lot of documentation on how to use NVIDIA but nothing on AMD. In my case I was only trying to mount /dev/dri into workloads containers as I assume it is enough.

How do you do that ? I mounted /dev/dri using a hostPath mount on the container but the scheduler ignored it (which is good). I am assuming the new securityContext rules mean it is not possible with the latest versions of k8s. How do I mount /dev/dri in a container in Talos 1.7.6/1.8.0 on kubernetes 1.31.1 without giving privileges to the container ?

I also tried

machine:
  kubelet:
    extraMounts:
    - destination: /dev/dri
      type: bind
      source: /dev/dri
      options:
        - bind
        - rshared
        - rw
  udev:
    rules:
    - SUBSYSTEM=="drm", KERNEL=="card*", GROUP="44", MODE="0660"
    - SUBSYSTEM=="drm", KERNEL=="renderD*", GROUP="44", MODE="0660"

but it didnt change anything so I am probably missing something ?

Thanks
b.

[smira]

Converted to this to be a discussion, as it's not an issue.

Please see https://www.talos.dev/v1.8/kubernetes-guides/configuration/device-plugins/, this is a way in general in Kubernetes to expose host devices to the workloads.

The patch you posted shouldn't be needed.