Native nvidia support

[joedborg]

Seeing https://github.com/NVIDIA/nvidia-docker#quickstart

Note that with the release of Docker 19.03, usage of nvidia-docker2 packages are deprecated since NVIDIA GPUs are now natively supported as devices in the Docker runtime.

I was wondering if this is coming to containerd, assuming this is an implementation in runc? Or will we continue to need to add nvidia-container-runtime to the config.toml?

[cpuguy83]

For vanilla containerd (not the cri plugin), the API is focused around the OCI spec.
So it's just about adding the right devices and hooks.
There is a package to assist with this: https://github.com/containerd/containerd/blob/master/contrib/nvidia/nvidia.go

[joedborg]

Thanks @cpuguy83, are there any docs with how to use that?

[cpuguy83]

WithGPUs is an oci.SpecOpts.
You can pass this into WithNewSpec: https://godoc.org/github.com/containerd/containerd#WithNewSpec
This is done when creating a container.

So something like:

client.NewContainer(ctx, "myContainer", containerd.WithNewSpec(containerd.WithDefaultSpec, nvidia.WithGPUs(...))

[crosbymichael]

I'm also looking into better ways to handle this in CRI.

But right now, I tested this the other day on my ubuntu system and the WithGPUs works great still. You can also use ctr run --gpus 0 ... to test this.