#Verification

Verify that Alauda GPU Management is installed and that GPUs are schedulable.

#1. Operator and operands are ready

Confirm the operator controller and the ClusterPolicy-managed components are running:

kubectl get clusterpolicy
kubectl get pods -A -l app.kubernetes.io/managed-by=gpu-operator

The ClusterPolicy should reach a ready state, and the driver (unless in preinstalled mode), container toolkit, device plugin, GPU Feature Discovery, and DCGM-Exporter Pods should be running on the GPU nodes.

#2. Nodes advertise GPU resources

Check that GPU nodes report allocatable nvidia.com/gpu:

kubectl get node ${nodeName} -o jsonpath='{.status.allocatable}'
# The output contains "nvidia.com/gpu":"1" (the value depends on the number of GPUs)

If the value is missing, confirm the driver is healthy for the selected driver mode and that the device plugin Pod is running on the node. Do not assume a resource key; confirm it from node allocatable resources as shown.

#3. A workload sees the GPU

Deploy a short GPU workload and confirm it can access the device:

kubectl run -it --rm gpu-verify \
  --restart=Never \
  --image=<your-cuda-image> \
  --limits=nvidia.com/gpu=1 \
  -- nvidia-smi -L

The command should list the GPU. When deploying GPU applications, request the GPU as a resource limit:

spec:
  containers:
    - name: gpu
      image: your-image
      resources:
        limits:
          nvidia.com/gpu: 1   # request 1 GPU (required)

With the container toolkit configured, the nvidia RuntimeClass is applied so the container can reach the device. For the full workload scenario and ACP quota handling, see Request GPU resources.

#4. Metrics (optional)

If DCGM-Exporter is enabled, verify GPU metrics and dashboards as described in DCGM-Exporter.