Alauda Build of HAMi
  • English

    • Verification

    This page describes how to verify the HAMi product combination required by the selected deployment scenario. Verify HAMi-WebUI only when it is installed.

    Use the current delivery package and version page as the source of truth for backend support and resource key names. The NVIDIA examples below use the common HAMi NVIDIA resource keys. Ascend backend keys are package-dependent.

    TOC

    Verify HAMi coreVerify the NVIDIA backendVerify the Ascend backendVerify ACP quota metadataVerify workload schedulingVerify monitoringNext steps

    Verify HAMi core

    Check whether the HAMi scheduler is running:

    kubectl get pods -n kube-system | grep "hami-scheduler"

    Verify the NVIDIA backend

    For HAMi on NVIDIA GPU, check the built-in NVIDIA device plugin:

    kubectl get pods -n kube-system | grep "hami-device-plugin"

    Check whether the NVIDIA node reports HAMi resources:

    kubectl get node ${nodeName} -o=jsonpath='{.status.allocatable}'

    The output should contain resource names such as nvidia.com/gpualloc. If the node also reports direct vendor resource keys for the same GPUs, verify which device plugin currently owns device exposure.

    Verify the Ascend backend

    For HAMi on Ascend NPU or Ascend vNPU, check the OLM state, custom resource, DaemonSet, and active owner for Alauda Build of HAMi Ascend Device Plugin:

    kubectl get subscription,csv,installplan -A \
  | grep hami-ascend-device-plugin
kubectl -n <hami-namespace> get hamiadp hami-ascend-device-plugin -o yaml
kubectl -n <hami-namespace> get ds hami-ascend-device-plugin -o wide
kubectl get ds -A -o wide \
  | grep -E 'ascend-device-plugin|hami-ascend-device-plugin'

    The CSV must be Succeeded, the custom resource must be reconciled without an error, and the DaemonSet desired and ready counts must match. The native Ascend Device Plugin must be stopped on the target nodes. If ownership has not converged, follow Switch the Ascend Device Plugin Owner before continuing.

    Check whether the Ascend node reports the resource names documented by the installed HAMi and HAMi Ascend Device Plugin package combination:

    kubectl get node ${nodeName} -o=jsonpath='{.status.allocatable}'

    If the expected HAMi-managed resources are missing, verify whether the native Ascend Device Plugin is still running.

    Verify ACP quota metadata

    If users manage accelerator quotas in ACP, confirm that ACP quota metadata has been registered for the same HAMi resource keys that appear in node allocatable resources. Node allocatable resources and ACP quota fields are separate checks.

    Verify workload scheduling

    Deploy a workload that requests HAMi resources. For the NVIDIA backend example:

    spec:
  containers:
    - image: your-image
      imagePullPolicy: IfNotPresent
      name: gpu
      resources:
        limits:
          cpu: '2'
          memory: 4Gi
          nvidia.com/gpualloc: 1     # Request 1 physical GPU (required)
          nvidia.com/gpucores: "50"  # Request 50% of the compute resources per GPU (optional)
          nvidia.com/gpumem: 8000    # Request 8000MB of video memory per GPU (optional)

    If the workload is scheduled successfully, check the accelerator usage from the node, ACP monitoring, or HAMi WebUI according to your installed monitoring path.

    For an Ascend scenario, deploy a representative workload using the resource key documented by the installed HAMi and HAMi Ascend Device Plugin package combination. Confirm that the Pod uses runtimeClassName: ascend, reaches Running, and can run npu-smi info plus a device-memory or inference workload. Delete the Pod and confirm that the device is released. Use HAMi on Ascend NPU for whole-card semantics or HAMi on Ascend vNPU for slicing semantics.

    For NVIDIA backend nodes, you can also check process-level usage on the node:

    nvidia-smi pmon -s u -d 1

    Verify monitoring

    After workloads have been running for a while, open Administrator -> Operations Center -> Monitor -> Dashboards and check the HAMi dashboard.

    If HAMi-WebUI is installed, open {platform-url}/clusters/{cluster-name}/hami-webui and check whether resource overview data is displayed. When NodePort is enabled, also verify the backend query path:

    hami_webui_url=http://<node-ip>:<node-port>

curl --silent --show-error --fail \
  -H 'Content-Type: application/json' \
  --data '{"query":"up"}' \
  "${hami_webui_url}/api/vgpu/v1/monitor/query/instant-vector"

    The WebUI query must return HTTP 200 and a JSON object with a data array. A page that loads while this request returns HTTP 523 has not passed monitoring verification.

    For NVIDIA backend metrics, verify the associated DCGM-Exporter package according to the NVIDIA GPU metrics documentation when that metrics path is part of the deployment. For Ascend backend metrics, verify NPU Exporter according to the Ascend NPU monitoring documentation when the exporter component is enabled in the NPUOperatorCtl instance.

    Next steps