Use runtime nvidia

Hi all,

I’m trying to use docker spawner with runtime nvidia set in order attach GPUs installed on my JUPYTERHUB server.

I tried with the following snippet in my jupyterhub_config.py file but I was unlucky until now.

c.DockerSpawner.extra_create_kwargs.update({ 'runtime': 'nvidia' })

Can you help me understand how to set it up properly?

Thanks.

[minrk]

Thanks for the working example! I’m not sure what a PR would include, though. From the looks of what you have there, it seems like extra_host_config is already enough:

import docker

c.DockerSpawner.extra_host_config = {
    "device_requests": [
        docker.types.DeviceRequest(
            count=-1,
            capabilities=[["gpu"]],
        ),
    ],
}

which seems to work when I try it, so we already have this feature satisfied by passing through arguments to docker-py.

[minrk]

Following up on that, the extra_create_kwargs is unnecessary. Only the extra_host_config is required. The main source of confusion was that the docker Python docs suggest that you can specify runtime via create, but this is only true if there is not a host_config also specified. Since DockerSpawner always specifies a host config, runtime can only be specified there.