Kubeflow

Kubeflow is an open-source machine learning (ML) platform built on Kubernetes. It provides a suite of tools and frameworks to develop, deploy, and manage ML workflows at scale. One of these tools is KServe, which is used to support inference endpoints. EGS also uses KServe to manage inference endpoints. By default, EGS installs a version of KServe to support EGS-driven inference endpoints. Kubeflow includes its own version of KServe. Therefore, if you want to use KubeFlow with EGS, you must disable KServe in the EGS installer script and configure EGS to use Kubeflow's KServe.

To use Kubeflow with EGS:

1. Disable KServe in the EGS installer YAML script.

2. Install EGS.

3. Install Kubeflow.

4. Modify the default Kubeflow deployment mode to Raw.

5. Modify the ingress in Kubeflow to use nginx instead of the default istio.

   The following part of the installer script snippet shows the required changes:

   kserve:
    enabled: false
    kserve:
      controller:
        gateway:
          domain: ""
          ingressGateway:
            className: nginx

6. Create Inference Endpoints on the EGS Admin/User Portal and ensure they are created in the Kubeflow's kserve namespace. For more information, see Manage Inference Endpoints.