Helm Chart Deploys NVIDIA Run:ai and KAI Scheduler on Kubernetes

Summary

The logicalclocks/run-ai-helm repository provides a Helm chart for deploying NVIDIA's open-source Run:ai 3.x line on Kubernetes. It wraps the upstream control-plane and cluster charts, KAI Scheduler components, kube-prometheus-stack, and NVIDIA or fake GPU operators, adding release profiles, validation, required Secrets, and cluster registration. The recommended installation uses separate releases for prerequisites, the Run:ai control plane in the runai-backend namespace, and the Run:ai cluster in the runai namespace; the control plane and cluster cannot share a release. Cluster installation depends on a running control plane because pre-install hooks use a cluster UUID and bootstrap token, and the registration script must be rerun before each cluster install or upgrade because the control plane mints a new token each time. The chart requires Kubernetes 1.30 or later, Helm, kubectl, an ingress controller, DNS and TLS for the control-plane domain, persistent storage, Prometheus monitoring, and an NVIDIA GPU stack. A development mode replaces the NVIDIA GPU Operator with a fake operator that reports eight Tesla-A10 GPUs per labelled node. KAI Scheduler can also be installed independently, but that profile cannot be rendered together with the Run:ai cluster profile. OpenShift requires an explicit values file for Routes and SCCs, while ArgoCD users must set install or upgrade mode because Helm template cannot resolve lookups and should create one Application per release in dependency order. The repository targets the open-source 3.x charts and distinguishes them from NVIDIA's commercial 2.x charts, which use an NGC API key and proprietary runai-engine scheduling.