As part of the initiative to expand Red Hat OpenShift AI (RHOAI) and the Distributed Inference with llm-d product to non-OpenShift Kubernetes environments (xKS), such as AKS, EKS, and CoreWeave, we are adopting Helm as the primary packaging and installation standard.

While llm-d is the initial target for this rollout, the roadmap includes expanding support to other RHOAI components in the future. The RHOAI Operator serves as the unified delivery vehicle to bring these components and features to non-OpenShift Kubernetes distributions, ensuring a consistent operational experience across all supported environments.

The RHOAI Operator relies on specific OpenShift operators for distributed workload orchestration, specifically:

• LeaderWorkerSet

• Kueue

• JobSet

On OpenShift, these components are typically managed via OLM (Operator Lifecycle Manager). However, on target xKS environments, OLM is often not present, is not the preferred method for managing infrastructure components, or is unsupported. Helm has been selected as the standard package management solution for provisioning RHOAI and its associated dependencies on non-OpenShift Kubernetes platforms.

Problem Statement

Currently, there is no official "Red Hat supported" path to install the OpenShift JObs Operator using Helm.

Although it would be possible to install the upstream JobSet controllers using community Helm charts, we explicitly aim to preserve the same semantics and operational logic by leveraging the related OpenShift Operators. This ensures the platform functions consistently across vendors and cloud providers, rather than managing divergent behaviors between "upstream" and "productized" versions.

Requirements

1. Develop Helm charts for the JobSet operators.

1. Publish these charts as OCI artifacts to the Red Hat Registry (registry.redhat.com)

1. Ensure the charts and the underlying operators are validated to run on standard Kubernetes distributions without hard dependencies on OpenShift specific APIs (or with those dependencies made optional).