Prereqs: vendored cluster-level GPU artifacts

Manifests that must exist on the cluster before the rest of ../platform.sh's layers come up — the device plugin advertises nvidia.com/gpu to the scheduler, and the exporter feeds GPU telemetry to Prometheus (what netfix's hostNetwork patch targets). install.sh applies them, wired into the prereqs layer / make -C bootstrap prereqs; see ../README.md's layer table.

Vendored as committed YAML instead of kubectl apply -f <upstream-url> / a live helm install (which is what this layer, and phase-2-capstone/gpu-node/diy-cluster.md, used to do) so the exact manifests applied to this cluster are reviewable, diffable, and pinned in git rather than resolved from a moving upstream tag/chart at apply time.

Files

File	What	Source
nvidia-device-plugin.yaml	NVIDIA k8s-device-plugin DaemonSet — advertises nvidia.com/gpu, one whole GPU per replica	NVIDIA/k8s-device-plugin v0.16.2, vendored + runtimeClassName: nvidia baked in
nvidia-device-plugin-timeslicing.yaml	Same DaemonSet plus a time-slicing ConfigMap — advertises 2 schedulable nvidia.com/gpu units per physical card	Same base, extended per the plugin's "Shared Access to GPUs" config
nvidia-gpu-exporter.yaml	GPU telemetry (util/mem/temp/power) on :9835/metrics, hostNetwork	plain-manifest port of the utkuozdemir/nvidia_gpu_exporter Helm chart's DaemonSet template, pinned to release v1.14.0 (image tag 1.14.0)

Apply

Via install.sh (what ../platform.sh prereqs / make -C bootstrap prereqs runs):

KUBECTL="k3s kubectl" ./install.sh up      # plain device plugin (default) + exporter
GPU_TIME_SLICING=1 KUBECTL="k3s kubectl" ./install.sh up   # time-sliced variant instead
./install.sh down                          # tear both down

Or by hand — pick one of the two device-plugin manifests, they share a name/namespace so applying one replaces the other:

# Plain: one nvidia.com/gpu per physical card (default; matches vllm-2gpu.yaml — one
# replica per GPU, see phase-2-capstone/gpu-node/vllm-2gpu.yaml)
kubectl apply -f nvidia-device-plugin.yaml

# OR time-sliced: 2 schedulable nvidia.com/gpu units per physical card, to demonstrate
# a shared card (see phase-2-capstone/gpu-node/vllm-timeslice.yaml)
kubectl apply -f nvidia-device-plugin-timeslicing.yaml

# Either way, then the exporter:
kubectl apply -f nvidia-gpu-exporter.yaml

Re-running the device plugin's DaemonSet pods restarts every pod currently using a GPU (the plugin re-registers its resource pool), so switch variants between demo runs, not mid-run.

Why time-slicing, not MIG

MIG (Multi-Instance GPU) partitions a card into hardware-isolated slices, but it's an Ampere-and-later datacenter feature (A100, H100, ...) — nvidia-smi mig -lgip reports unsupported on GeForce/RTX, full stop, no driver version fixes it. Neither of this cluster's cards (RTX 3060 Ti, RTX 4070 Laptop — see diy-cluster.md) can do it. The applicable mechanism on consumer GeForce hardware is CUDA time-slicing (what's configured here — interleaved compute, no isolation between the sharing workloads) or MPS (space-partitioned, needs a control daemon, mutually exclusive with time-slicing). Time-slicing is also what makes maxReplicaCount: 2 in keda-scaledobject-gpu.yaml worth revisiting past "one replica per card" — see vllm-timeslice.yaml.