[ANNOUNCE] Kubernetes-native API for Hyperparameter Optimization at Scale

[Andrey Velichkevich]

Hi Folks,

We’re currently working on the next evolution of the Katib project, focused on building a Kubernetes-native API for hyperparameter tuning.
The new API will feature deep integration with TrainJob and the Kubeflow SDK, enabling optimization of LLMs and other large-scale models.

If you’re an existing Katib user, your input is especially valuable. As part of this effort, we expect to deprecate some existing functionality to make the API more robust and enterprise-ready.

Over the coming weeks, we plan to prepare an official KEP to redesign the Katib APIs. In the meantime, we’d greatly appreciate your feedback on the proposal below.

• Proposal: https://docs.google.com/document/d/1Y8IJ-UdZ7VCEAlax_xEFbbqEi7EB6SfIX4D7ua-xn4M/edit?tab=t.0#heading=h.bbjge5qtwrks
• Tracking issue: https://github.com/kubeflow/katib/issues/2605

Regards,

Andrey