k3d fails to pull custom gpu image

[kkwekkeboom]

Hi all,

I was following the guide on running CUDA workloads here. When creating a cluster using k3d cluster create gputest --image='<my-private-registry>/rancher/k3s:v1.21.2-k3s1-cuda' I get

Failed Cluster Creation: Failed to check for the existence of edac folder: Error response from daemon: Head https://<my-private-registry>/v2/rancher/k3s/manifests/v1.21.2-k3s1-cuda: unauthorized: authentication required

This should not be the case as docker pull <my-private-registry>/rancher/k3s:v1.21.2-k3s1-cuda works fine on the same host.

Providing a registries.yaml does not seem to help. Any suggestions?

[iwilltry42]

Hi @kkwekkeboom, thanks for starting this discussion!

Providing a registries.yaml does not seem to help.

The error is really about k3d (not k3s in the container) failing to pull your image.
The problem here is, that docker doesn't provide an easy-to-use function in its SDK to use the Auth info (just checked the Docker CLI code and even there it's super complex), so I fear it won't easily land in k3d.
However, since you ran the docker pull, the image should be present and shouldn't be pulled anymore by k3d.. that's a thing we can fix and I did that here: db342a8
It will land in the next k3d release 👍