Allow selection of different GPU model types during notebook creation

Checks

• I have searched the existing issues.
• My request is related to one of the components in the kubeflow/notebooks repository.

Motivation

At the moment, GPUs can only be configured via the resource limit field. This works if the platform administrators want to expose full NVIDIA or AMD gpus via nvidia.com/gpu or amd.com/gpu or partitions via e.g. nvidia.com/12gb.1gb. If multiple GPU model types are offered, there is no easy way to request them from the user perspective. This is only possible via affinities which are not very user friendly.

Implementation

I would allow the administrators to set the node selector via the notebook configuration, additionally to the resource limit field.

For the users , it could look like this:

For the administrators, it the config could look like this:

flavors:
        - limitsKey: nvidia.com/gpu
          uiName: "NVIDIA H100 NVL (94GB)"
          id: "nvidia-h100-nvl-full"
          nodeSelector:
            "nvidia.com/gpu.product": "NVIDIA-H100-NVL"

We have a working version offering a flavors section next to existing vendors section which we are happy to contribute.

Are you willing & able to help?

• I am able to submit a PR!
• I can help test the feature!

Start by tracing notebook creation and the existing vendors and resource-limit configuration. Compare the working flavors approach described in the issue, including administrator-defined node selectors and the user-facing GPU choice. Done means administrators can configure GPU model options and users can select one that applies the corresponding node selector without breaking current GPU requests.