How to run kata gpu containers with k8s cluster

[lizhixin3016]

To run gpu containers with kata-runtime, we need to enable vfio driver for gpu devices,
but the k8s-device-plugin need to install nvidia driver on k8s node to detect the gpu
devices, causing we can not passthrough gpu devices by vfio-pci driver into the kata
containers.

I didn't find any introduction or document about this, anyone knows how to deal with it?

[Jimmy-Xu]

@lizhixin3016
Have you tried this device plugin https://github.com/NVIDIA/kubevirt-gpu-device-plugin?
It does not need to install nvidia driver on k8s node, and you can passthrough the VFIO device of Nvidia to the kata container.
You can install and load the Nvidia driver into the guest kernel rootfs image or kata container.

[lizhixin3016]

@Jimmy-Xu
This plugin looks useful to me, I will try deploy and run it with kata in k8s cluster.

Thank you very much!

[liubin]

@lizhixin3016 , if Jimmy's comment helped you, consider closing this issue?

[ariel-adam]

Closing this issue.