Helm and Kubernetes

Why it matters

A production NIM runs on a cluster, where image access, model cache storage and GPU limits become chart values that someone has to own.

What to practise

Fetch the nim-llm chart from helm.ngc.nvidia.com, read it with helm show readme and helm show values, and write a minimal values.yaml: image.repository and image.tag, persistence for the model cache, and a nvidia.com/gpu limit under resources. Leave model.ngcAPISecret empty for a keyless public image, and add NGC secrets only when the image or model requires a key. Install it with helm install, check pods and services with kubectl, port-forward, and call /v1/health/ready. Note how customArgs passes backend flags without replacing the image’s NIM_PASSTHROUGH_ARGS, and why a cluster without a spare GPU sets statefulSet.enabled: false and a Recreate deployment strategy.

How it fits

This is the Helm half of 7.2.3’s choice between the chart and the NIM Operator. The readiness check at the end repeats 7.2.1’s validation inside the cluster.