[Feature Request] Allow to configure CDI annotation prefix

[xiongzubiao]

This is to better support non-nvidia OCI runtimes in CDI mode. The gpu-operator uses nvidia.cdi.k8s.io as the default CDI annotation prefix. However, it requires to use Nvidia's container runtime. When using a non-nvidia OCI runtime, CRI-O or Containerd wants cdi.k8s.io prefix. Although one can edit the environment variables in the nvidia-container-runtime-daemonset and the nvidia-device-plugin-daemonset to achieve that, it is much easier that if the gpu-operator provides a chart config such as cdi.prefix.

[github-actions]

This issue is stale because it has been open 90 days with no activity. This issue will be closed in 30 days unless new comments are made or the stale label is removed. To skip these checks, apply the "lifecycle/frozen" label.

[cdesiniotis]

Hi @xiongzubiao, starting with GPU Operator 25.10.0 we use native-CDI support in runtimes (containerd, cri-o) by default. This means we are no longer using the nvidia runtime to inject GPU support into workload containers. Given this, I no longer think we need an option to configure the CDI annotation prefix. I encourage you to try out the latest release. Please re-open this issue if this is still relevant.