I want to have a container inside a pod, that uses some docker runtime (we can specify it using --runtime flag in pure docker).

In my case, I want to use nvidia-docker2 runtime (https://github.com/NVIDIA/nvidia-docker) for using GPU inside my containers in Kubernetes cluster.

Answer: Specify default runtime in /etc/docker/daemon.json like that:

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