Skip to content

Maintain Accelerator Models ​

Target Outcome ​

Define the NPU model once with the correct vendor, model identifier, memory, and resource key so specifications and monitoring use the same identity.

Applicable Roles ​

  • Platform Operator

Before You Start ​

  • Record the exact model and per-card memory reported by the four physical NPU cards.
  • Confirm that no existing entry already represents the same model.

Entry ​

  • Role: Operator
  • Menu: AI Infrastructure > On-Prem > Resource Pools > Accelerator Management
  • Route: /powerone/resourcepool/accelerators

Steps ​

  1. Filter by the target NPU vendor.

  2. Find the actual model and verify its family, memory, and adaptation state.

  3. If it is missing, create the accelerator and enter the Kubernetes resource name.

  4. Associate the model with the metric used by the cluster device plugin.

  5. Save and verify the expected managed state.

  6. Before using the model in specifications or inference templates, confirm that its vendor, model identifier, memory, and resource key match the values reported by the cluster device plug-in.

Accelerator management

Four-NPU Notes ​

  • This page defines the card model, not the installed quantity. The cluster reports the actual card count.
  • All four cards of the same model should use the same resource key.
  • Memory and model metadata affect flavor and deployment feasibility checks.

Completion Checklist ​

Purpose: These are the exit criteria for the current feature task. Use them to decide whether the result is observable and reviewable and whether you can continue to the next step in the scenario. They do not repeat the procedure; if any item fails, follow the troubleshooting section below.

CheckPass Criteria
1The NPU model is visible.
2Its adaptation state and metric association are correct.
3The resource key matches the cluster.

Troubleshooting ​

SymptomCheck First
The model is unavailable in a specificationAccelerator status, model identifier, and vendor mapping
Monitoring cannot match the deviceWhether the cluster-reported model matches this entry exactly
A downstream specification cannot use the acceleratorManaged/adaptation state, metric association, resource key, and cluster visibility

User Manual ​

Accelerator Management