Lab: a GPU-style cluster on your laptop
Create a local kind cluster, advertise a fake nvidia.com/gpu extended resource, taint a 'GPU' node, deploy a CPU fake-LLM through the production manifest pattern, then break it on purpose: Pending pods, probe loops, a stuck rollout, a PDB-blocked drain and a priority preemption.
Goal: practise the Kubernetes mechanics from this track on a cluster you can break safely: GPU-style scheduling, probes for slow starts, rollout strategy under a full pool, disruption budgets and priority preemption.
You need: Docker, kubectl, and kind (Kubernetes in Docker), about 4 GB of free RAM, and an hour. No GPU is required: you will advertise a fake nvidia.com/gpu resource and serve a CPU stand-in for an LLM.
Honest limits. The manifests and steps below follow the documented behaviour of Kubernetes and kind, but they have not been executed by the author of this lesson in every environment; if a command's output differs on your version, trust the live output and the Kubernetes documentation. This lab tests scheduling and orchestration, not CUDA, drivers, VRAM or NCCL. Those need real hardware and the verification checklist from GPUs on Kubernetes.
Pro content
Keep reading with Inference Academy Pro
Unlock this full lab, plus every advanced lesson, hands-on lab, the larger playground models and completion certificates.