taugrid/examples/aks-gpu-quickstart/nvidia-device-plugin.yaml at main · Azure/taugrid · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
# NVIDIA device plugin DaemonSet.
#
# WHY THIS FILE EXISTS
# --------------------
# `az aks nodepool add` with a GPU SKU installs the NVIDIA *driver* (the node
# pool reports gpuProfile.driver=Install and the node is labelled
# accelerator=nvidia), but on current AKS it does NOT deploy a device plugin.
# Without one the kubelet never advertises the nvidia.com/gpu extended
# resource, so Kueue cannot admit a GPU workload and the pod sits Pending on
# "Insufficient nvidia.com/gpu" with nothing in the events to explain it.
#
# This is cluster infrastructure, not TauGrid configuration: the TauGrid chart
# is unchanged and every install still looks the same.
#
# IMAGE SOURCING
# --------------
# Pinned by digest to the Microsoft-mirrored copy on MCR rather than to
# nvcr.io, per AGENTS.md > "Container image sourcing": never reference an
# external mutable tag from a manifest that runs on a cluster, and this one is
# privileged with host device access.
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:
      # Tolerate a tainted GPU pool so this still works if an operator chose to
      # taint theirs. The quickstart's own pool is untainted because TauGrid's
      # stock ResourceFlavor declares no tolerations.
      tolerations:
        - key: nvidia.com/gpu
          operator: Exists
          effect: NoSchedule
        - key: sku
          operator: Equal
          value: gpu
          effect: NoSchedule
      priorityClassName: system-node-critical
      # Only land on nodes AKS has actually labelled as GPU nodes; otherwise the
      # plugin crash-loops on every CPU node in the cluster.
      nodeSelector:
        accelerator: nvidia
      containers:
        - name: nvidia-device-plugin-ctr
          image: mcr.microsoft.com/oss/v2/nvidia/k8s-device-plugin@sha256:78410ff8b27e623df1fa53e65a81fe57bdff26d2c2e88decf58d8bd08323738d
          env:
            # AKS sets containerd's default runtime to nvidia-container-runtime,
            # but that hook only injects the GPU devices and driver libraries
            # into a container when NVIDIA_VISIBLE_DEVICES is set. Without these
            # two vars the plugin loads NVML successfully and then reports
            # "No devices found. Waiting indefinitely." -- it can see the
            # library but not the hardware.
            - name: NVIDIA_VISIBLE_DEVICES
              value: "all"
            - name: NVIDIA_DRIVER_CAPABILITIES
              value: "all"
            - name: FAIL_ON_INIT_ERROR
              value: "false"
          securityContext:
            # Privileged plus an explicit /dev mount, rather than relying on the
            # nvidia-container-runtime hook to inject the devices. The hook is
            # configured on AKS but did not inject them for this pod even with
            # NVIDIA_VISIBLE_DEVICES set, and a plugin that cannot enumerate
            # devices fails silently ("No devices found") instead of crashing.
            privileged: true
          volumeMounts:
            - name: device-plugin
              mountPath: /var/lib/kubelet/device-plugins
            - name: dev
              mountPath: /dev
      volumes:
        - name: device-plugin
          hostPath:
            path: /var/lib/kubelet/device-plugins
        - name: dev
          hostPath:
            path: /dev