Alauda Ascend NPU

Prerequisites

Prepare the cluster and nodes before installing Alauda Build of NPU Operator.

TOC

ACP and cluster accessNode and hardwareDriver and runtimeDevice exposure modelOptional componentsACP cross references

ACP and cluster access

  • ACP v4.0-v4.3.
  • Cluster administrator access to the target workload cluster where Alauda Build of NPU Operator will be installed.
  • OperatorHub and Cluster Plugin installation are available in ACP.
  • Package download from Customer Portal and package upload in ACP are available.

Node and hardware

  • Ascend 910B or Ascend 310P worker nodes.
  • ARM architecture.
  • Alauda Build of Node Feature Discovery is installed on the target workload cluster so shared kernel, OS, architecture, and hardware labels are available before NPU components are reconciled.
  • Nodes selected for NPU components allow NPU Feature Discovery to run, or you have identified nodes that require manual placement-label fallback. Verify masterselector=dls-master-node and workerselector=dls-worker-node after creating the NPUOperatorCtl instance; they are normally maintained automatically rather than added as an installation prerequisite.
  • For Ascend 910-family nodes, complete HCCN device IP configuration before workload verification. v26.6.0 does not initialize HCCN device IPs automatically.

Driver and runtime

Choose one driver lifecycle mode before creating the NPUOperatorCtl instance:

ModeWhen to use
Pre-installed driverUse when the host driver is installed and upgraded outside Alauda Build of NPU Operator. Disable Driver management in NPUOperatorCtl.
Operator-managed ordinary OS installUse when the host OS can run the delivered driver installer flow. Enable Driver management and keep precompiled driver delivery disabled.
Operator-managed immutable OS / Alauda OS precompiled installUse when the host OS is immutable or should use a precompiled driver tree. Enable Driver management and enable precompiled driver delivery.

In pre-installed driver mode, verify that /usr/local/Ascend/driver exists on every selected NPU node before enabling NPU components that depend on the driver.

For operator-managed immutable OS / Alauda OS precompiled mode, import the matching ARM64 Driver image separately from the product package. Map the complete source image reference to the target workload cluster registry and make sure the admitted image is allowed by ImageWhiteList when the cluster enforces image policies.

Device exposure model

Choose one Ascend device exposure owner for each node:

ModelDevice pluginWhen to use
NPU Operator-managed direct NPU allocationNPU Operator Ascend Device PluginUse when Alauda Build of NPU Operator owns node lifecycle and workloads request NPUs directly.
Manually managed direct NPU allocationHuawei Ascend Device PluginUse when you do not use NPU Operator for full lifecycle management but workloads still request NPUs directly. Use the Huawei MindCluster documentation as the upstream reference.
HAMi-managed allocation or vNPU behaviorHAMi Ascend Device PluginUse when workloads need HAMi scheduling, virtualization, or slicing behavior on Ascend devices.

Do not run multiple device exposure owners for the same Ascend devices on the same node unless the current product versions explicitly document and validate that topology.

Optional components

  • Install Volcano only if you plan to enable ClusterD.
  • Pre-stage MindIO SDK zip files on each target NPU node only when you plan to enable MindIO TFT or MindIO ACP in an offline or controlled network environment.

ACP cross references