2.5 Checkpoint

Time: 5 min buffer + 10 min break

Run the validator. It checks everything Lab 1 was supposed to produce.

./scripts/validate-lab1.sh

Lab 1 validation
  ✅ 1 GPU node(s) from NodePool 'gpu'
  ✅ allocatable nvidia.com/gpu on GPU node(s): 1
  ✅ nvidia-device-plugin-daemonset ready on 1 node(s)
  ✅ nvidia-dcgm-exporter ready on 1 node(s)
  ✅ gpu-feature-discovery ready on 1 node(s)
  ✅ no nvidia-driver-daemonset (driver.enabled=false — correct on EKS)
  ✅ gpu-smoke ran nvidia-smi: NVIDIA L4
🎉 Lab 1 checkpoint reached.

Checkpoint 1 — show your nvidia-smi

Post your nvidia-smi screenshot (or the validator output) in the chat, or hold your laptop up if we're in a room. GPU model and driver version visible. Helpers are sweeping for anyone with a ❌.

If something is ❌

Symptom	First thing to check
No nodes from the NodePool after 5 min	kubectl describe nodeclaim — look at events. Quota (VcpuLimitExceeded) and InsufficientInstanceCapacity across all four types are the usual suspects. See Troubleshooting → Karpenter.
Node exists but nvidia.com/gpu never appears	kubectl -n gpu-operator get pods -o wide — is the device plugin on the node? Is nvidia-driver-daemonset there when it shouldn't be?
gpu-smoke stuck in ContainerCreating	kubectl describe pod gpu-smoke — image pull vs. device allocation.
nvidia-smi prints "couldn't communicate with the NVIDIA driver"	The container got scheduled but not injected. Almost always the toolkit/CDI mismatch — confirm cdi.enabled=false and toolkit.enabled=false in helm get values gpu-operator -n gpu-operator.

What you just built

Say it back to yourself; you'll be asked in Module 3:

• Who put the driver on the node? (the AMI)
• Who made nvidia.com/gpu: 1 appear on the node? (the device plugin, deployed by the Operator, registered with kubelet)
• What stopped CoreDNS from landing on your $1/hour node? (the taint)
• Why did the pod wait ~3 minutes? (no node existed; Karpenter had to launch one and wait for it to initialise)

Leave things as they are

Don't delete the pod or the node. The node is now empty (the pod completed), and with consolidationPolicy: WhenEmpty Karpenter would reclaim it after consolidateAfter — which is why that's set to 2 hours for today. Module 3 puts the node back to work.

☕ Break — 10 minutes

We resume with Module 3 →