[FEATURE] Allow setting the container runtime for the k3d containers

[sin-ack]

Is your feature request related to a problem or a Pull Request

I'm trying to run some Vulkan workloads on a k3d cluster on a machine that has an NVIDIA GPU. To do this, Vulkan needs to have the /etc/vulkan/icd.d and /etc/vulkan/implicit_layer.d files passed in for NVIDIA. This is normally done by both nvidia-container-toolkit and gpu-operator, but is not done on its own with --gpus all. The "nvidia" runtime must be used when running the k3d container with Docker so that these files are exposed to the k3d nodes, so that the k3d nodes can re-expose them to the containers they run.

Scope of your request

A new flag/config option under runtime.

Describe the solution you'd like

An option like runtime.dockerRuntime on the Simple object that allows me to set which runtime I want to use for running the k3d containers.

Describe alternatives you've considered

Currently the following workaround exists:

• Create the cluster
• Stop the docker service
• Edit /var/lib/docker/containers/[k3d container id]/hostconfig.json
• Set "Runtime" to "nvidia"
• Start the docker service

But obviously this doesn't persist if I recreate the cluster and is therefore brittle.

[rassie]

In terms of workarounds, it should also be possible to set the default runtime for Docker to nvidia in /etc/docker/daemon.json:

{
    "default-runtime": "nvidia",
}

Otherwise: yes please, need the feature too.

[dleviminzi]

+1 for needing this feature.