Kubernetes Quickstart

The same images as the container deployment, with GPU wiring selected by one value.

Install

helm repo add lifeboat https://raw.githubusercontent.com/IterateAI/lifeboat-releases/main/helm
helm repo update
helm install lifeboat lifeboat/lifeboat \
  --set gpu.vendor=nvidia \
  --set admin.password=<a real password>

The two required decisions

gpu.vendor takes nvidia, amd or lite, and switches the image, the GPU resource, the runtime wiring and the vendor-specific pod settings together. Setting it wrong does not degrade — an AMD node given the NVIDIA wiring fails admission.

admin.password is required: the chart refuses to render without it, rather than deploying something with a default credential. Use admin.existingSecret to point at a Secret you manage instead.

Cluster prerequisites

Vendor	Needs
nvidia	NVIDIA device plugin or GPU Operator, driver 580.65.06+
amd	AMD GPU device plugin or GPU Operator
lite	nothing

Persistence

Give the deployment a volume for the data directory and one for models. Without the data volume you lose the registry, users, audit log and the licence — and the cluster identity a licence is bound to, which means the licence has to be re-issued.

Ingress and TLS

Front the control plane with your own ingress for TLS, hostname routing and access control. Lifeboat can also terminate TLS itself when given a certificate and key.

Shared memory

Tensor parallelism above 1 needs /dev/shm sized for it. The chart exposes shmSize; it is fixed at pod creation, so changing it recreates the pod. Getting it wrong surfaces as insufficient_shared_memory at server start, with the remedy named — see Server lifecycle.

Verify

kubectl get pods -l app.kubernetes.io/name=lifeboat
kubectl port-forward svc/lifeboat 8001:8001
curl -s localhost:8001/api/version

Sources and references

• Values and vendor switches: Deployment on Kubernetes
• Licence and cluster identity: Licensing