Use GPU with kubernetes yaml support in podman.

[andrew-kennedy]

I've successfully used gpu support in the docker-style container creation shell shell incantations within podman, for applications like plex and jellyfin running in containers, as I've installed the nvidia-container-runtime and with the specification of the proper runtime/gpu flags, everything works.

I'm trying to move my podman pods configuration into kubernetes yaml, for ease of eventual transition to k8s. Is there a good way to use GPUs when configuring containers to be run with podman play kube?

[rhatdan]

Would love to have you write an article on what you have done.
@umohnani8 @haircommander @mrunalp @saschagrunert Could you answer the question on kube?

[saschagrunert]

We could use the Kubernetes resource indicators to setup the correct runtime and its flags:

      resources:
        limits:
          nvidia.com/gpu: 1 # requesting 1 GPU

This should work in the same way for AMD and Intel graphics.

[fpoirotte]

See also #17833 (and #14934).

As pointed out by @saschagrunert, support for GPUs (or other types of devices) in Kubernetes v1.26+ uses the Device Plugins Framework: you install a vendor-supplied plugin on your cluster, then request the device as a resource in your YAML file through a vendor-specific keyword (similar to how annotations work).

Implementing the framework in podman may prove difficult if support for every existing vendor is required.
E.g. in my case, I would like to use the Kubernetes Generic device plugin which uses completely different keywords (by default, these are named squat.ai/video, squat.ai/serial, squat.ai/fuse, squat.ai/audio, squat.ai/capture, but the names can be customized and new keywords added as well).