Alauda Ascend NPU
  • English

    • Configure Workloads

    Configure NPU workload runtime behavior after you have selected the resource key. With the default admission webhook enabled, keep workload manifests focused on the NPU resource request and let the webhook add the ascend RuntimeClass.

    This page covers direct NPU allocation through Alauda Build of NPU Operator. For HAMi sharing, virtualization, vNPU behavior, or HAMi-specific resource keys on Ascend devices, use the HAMi documentation.

    TOC

    Resource requestRuntimeClass behaviorRuntime integration verification hintsScheduling and quota

    Resource request

    For the workload resource field and examples, see Request NPU Resources. Use the exact resource key reported by the target node. Common examples include huawei.com/Ascend910 for 910-family devices, huawei.com/Ascend310P, and other environment-specific huawei.com/Ascend* keys. Do not derive the key only from the chip product name.

    RuntimeClass behavior

    NPU Operator v26.6.0 uses the ascend RuntimeClass backed by ascend-docker-runtime. The runtime integration DaemonSet stages the ascend-docker-runtime payload and uses npu-container-toolkit to configure the corresponding containerd handler on each NPU node.

    The admission webhook adds the ascend RuntimeClass to Pods that request Ascend resources. New manifests do not need to set runtimeClassName when the webhook is enabled. If the webhook is disabled, set runtimeClassName: ascend explicitly.

    Keep operator.runtimeClass set to ascend in v26.6.0. The runtime integration DaemonSet registers the containerd handler with that name, so changing only the Operator setting creates a RuntimeClass-to-containerd handler mismatch.

    Runtime integration verification hints

    When troubleshooting a workload that requested an NPU resource but cannot see devices, verify that the admitted Pod uses the ascend RuntimeClass and that the runtime integration DaemonSet is ready. In precompiled Driver mode, Driver libraries are staged from the Host runtime Driver tree under /run/ascend/driver; in pre-installed Driver and ordinary OS managed install modes, components use /usr/local/Ascend/driver.

    Inside a workload container, check for paths such as /usr/local/bin/npu-smi, /usr/local/Ascend/..., and /dev/davinci*. Do not confuse these workload-side paths with the host-side driver tree.

    Scheduling and quota

    If a workload stays Pending:

    1. Check that nodes report the requested resource key.
    2. Check project and namespace quota in ACP.
    3. Check whether the node is in driver upgrade or recovery state.

    For quota field display in ACP, see Accelerator Resource Quota.