A Helm chart for Ollama that handles GPU envy and Knative quirks

What it does This community chart is essentially packaging glue that deploys Ollama onto Kubernetes, turning infrastructure boilerplate—GPU node scheduling, persistent storage, ingress, and autoscaling—into configurable values. You can set it to pull specific models at startup or create custom model variants from templates before the container serves traffic.

The interesting bit The chart’s Knative mode is a nice touch: it swaps the standard Deployment for a Knative Service, but because Knative pods are ephemeral, it renders a separate bootstrap Job to pre-load models. The Job is non-blocking and hash-named against your model list, so changing models on upgrade automatically retriggers bootstrap without stalling the release.

Key highlights

• Supports NVIDIA and AMD GPU scheduling out of the box, though the README notes AMD support in upstream Ollama itself is still spotty.
• Can pull models at container startup or create custom variants from Modelfile templates via values.yaml.
• Optional Knative Service deployment with a companion bootstrap Job for model persistence.
• Exposes standard Helm levers like autoscaling, affinity, and ingress hosts.

Caveats

• AMD GPU support is explicitly flagged as spotty in upstream Ollama.
• Knative mode requires enabling specific feature flags on your cluster for PVC mounts; otherwise the Service gets rejected by webhooks.
• The project is currently migrating its Helm registry URL, so check where you are pulling from.

Verdict Worth a look if you are already running a Kubernetes cluster and want a repeatable, version-controlled way to host Ollama without maintaining your own manifests. If you are just experimenting with local LLMs on a single machine, this much orchestration is probably overkill.