# GPU time-slicing variant of nvidia-device-plugin.yaml — applies INSTEAD of it (same
# DaemonSet name/namespace, so applying one replaces the other) to demonstrate two
# vLLM replicas sharing one physical 8 GB card.
#
# Why time-slicing and not MIG: MIG (Multi-Instance GPU) partitions a card into
# hardware-isolated slices, but it is an Ampere-and-later **datacenter** feature (A100,
# H100, ...) gated out of consumer silicon entirely — `nvidia-smi mig -lgip` reports
# "not supported" on a GeForce/RTX card, full stop, driver version is irrelevant. Neither
# of this cluster's cards (RTX 3060 Ti, RTX 4070 Laptop, see
# ../../phase-2-capstone/gpu-node/diy-cluster.md) can do it. The applicable mechanism on
# GeForce is CUDA **time-slicing** (interleaved compute, no isolation — a crashed replica
# can take the other down with it) or **MPS** (space-partitioned, needs a control daemon).
# Time-slicing is what's configured here: no daemon, no driver requirement beyond what's
# already installed, and it's exactly what makes `maxReplicaCount > 2` in
# ../../phase-2-capstone/gpu-node/keda-scaledobject-gpu.yaml meaningful on a two-GPU
# cluster — without it, `nvidia.com/gpu` capacity hard-caps replicas at one per card.
#
# Config format is upstream NVIDIA k8s-device-plugin v0.16.2 (see its README, "Shared
# Access to GPUs" / "With CUDA Time-Slicing"): the plugin is pointed at this ConfigMap via
# the `CONFIG_FILE` env var, mounted at /etc/nvidia/config.yaml.
apiVersion: v1
kind: ConfigMap
metadata:
  name: nvidia-device-plugin-timeslicing-config
  namespace: kube-system
data:
  config.yaml: |
    version: v1
    sharing:
      timeSlicing:
        resources:
          - name: nvidia.com/gpu
            # 2 replicas per physical GPU: enough to demonstrate a shared card
            # (2 vLLM pods on 1 GPU, see phase-2-capstone/gpu-node/vllm-timeslice.yaml)
            # without slicing an 8 GB card thin enough that neither replica boots.
            replicas: 2
---
apiVersion: apps/v1
kind: DaemonSet
metadata:
  name: nvidia-device-plugin-daemonset
  namespace: kube-system
spec:
  selector:
    matchLabels:
      name: nvidia-device-plugin-ds
  updateStrategy:
    type: RollingUpdate
  template:
    metadata:
      labels:
        name: nvidia-device-plugin-ds
    spec:
      runtimeClassName: nvidia
      tolerations:
        - key: nvidia.com/gpu
          operator: Exists
          effect: NoSchedule
      priorityClassName: "system-node-critical"
      containers:
        - image: nvcr.io/nvidia/k8s-device-plugin:v0.16.2
          name: nvidia-device-plugin-ctr
          env:
            - name: FAIL_ON_INIT_ERROR
              value: "false"
            - name: CONFIG_FILE
              value: /etc/nvidia/config.yaml
          securityContext:
            allowPrivilegeEscalation: false
            capabilities:
              drop: ["ALL"]
          volumeMounts:
            - name: device-plugin
              mountPath: /var/lib/kubelet/device-plugins
            - name: timeslicing-config
              mountPath: /etc/nvidia
      volumes:
        - name: device-plugin
          hostPath:
            path: /var/lib/kubelet/device-plugins
        - name: timeslicing-config
          configMap:
            name: nvidia-device-plugin-timeslicing-config