Added Kubeflow Pipeline component for ModelKit packaging and OCI push

[TheCoder2010-create]

Kubeflow Pipelines lacked a reusable component for packaging trained model artifacts as ModelKits and pushing them to OCI-compliant registries. This gap made it difficult to standardize model packaging and automate model delivery to registries within Kubeflow workflows.

What Has Been Added
Reusable Kubeflow Pipeline Component: A new component that:
Extracts training parameters, pipeline run ID, experiment name, and metrics from Kubeflow ML Metadata.
Packages model artifacts using the ModelKit format.
Pushes the packaged ModelKit to any OCI-compliant registry.
Supports authentication via Kubernetes secrets or IRSA/Workload Identity for cloud environments.
Component Implementation:
component.py: Python script for metadata extraction, packaging, and push logic.

Dockerfile: Container image definition with all dependencies.
component.yaml: KFP component specification for easy integration.
This addition enables seamless, standardized, and secure model packaging and registry integration for Kubeflow users.

[bmicklea]

Could you drop a Markdown file or other notes here that we can use to create docs for this? Or create a companion docs PR?

[TheCoder2010-create]

Could you drop a Markdown file or other notes here that we can use to create docs for this? Or could you create a companion docs PR?

KFP ModelKit Component

Overview

This Kubeflow Pipeline (KFP) component enables packaging a ModelKit model and pushing it to an OCI (Open Container Initiative) registry. It is designed to be used as a step in automated ML pipelines, making it easy to build, package, and distribute models using ModelKit within Kubeflow.

Files

• component.py: Python script implementing the component logic.
• component.yaml: KFP component specification for pipeline integration.
• Dockerfile: Container definition for the component runtime environment.

Usage

1. Build the Docker image:

   docker build -t <your-repo>/kfp-modelkit-component:latest .

2. Push the image to your registry:

   docker push <your-repo>/kfp-modelkit-component:latest

3. Use in a KFP pipeline:
   • Reference component.yaml in your pipeline definition.
   • Configure input parameters as needed (see below).

Inputs & Parameters

Refer to component.yaml for the full list of parameters. Typical parameters include:

• Model source path
• Model name and version
• OCI registry URL and credentials
• Additional packaging options

Example Pipeline Step

import kfp
from kfp.components import load_component_from_file

modelkit_op = load_component_from_file('component.yaml')

# Example usage in a pipeline
def my_pipeline(...):
    modelkit_op(
        model_path='path/to/model',
        model_name='my-model',
        version='1.0.0',
        registry_url='oci://my-registry',
        # ...other params...
    )

Notes

• Ensure all required environment variables and credentials are available at runtime.
• The component is designed for use with ModelKit and OCI-compliant registries.
• For advanced usage, customize component.py or the Dockerfile as needed.

---

[amisevsk]

Thanks @TheCoder2010-create for the contribution! This PR appears to be incomplete at the moment, but is a good starting point for implementing something like this. We still need to install the Kit CLI into the resulting container and use the correct sequence of commands to create a ModelKit.

You note in your PR description

Supports authentication via Kubernetes secrets or IRSA/Workload Identity for cloud environments.

However I don't see this referenced in the added files; is this something that's automatically done by the libraries you are using?

[amisevsk]

We can potentially reuse the standard KitOps image and copy the binary out of it into this image.