Skip to content

Scenario Overview - On-Prem Compute Onboarding ​

This scenario explains how to connect a local Kubernetes cluster and its GPU, NPU, or XPU devices to AGIOne, then expose the hardware through schedulable resource specifications. For the task “How do I manage four NPU cards?”, this scenario covers device identity, cluster onboarding, and flavor planning.

Applicable Roles ​

  • Platform Operator
  • Cluster administrator

Scenario Goals ​

  • AGIOne recognizes the target NPU model and resource key.
  • Cluster nodes report all four NPU cards.
  • Users can select one-card, two-card, and four-card resource specifications.

Scenario Flow ​

Main path: Create the region and availability zone → Identify the NPU model → Onboard the cluster and discover devices → Create one-, two-, and four-card plans → Verify downstream selection

StageKey Result
1. Establish the resource boundaryThe region and availability zone are available to the target cluster
2. Identify the modelNPU model, per-card memory, and scheduler resource key use one definition
3. Onboard the clusterCluster and nodes are available and all four physical NPU cards are discovered
4. Create plansOne-, two-, and four-card specifications match the real scheduling topology
5. Verify downstream useAn inference template or test workload can select and request the plan

Before You Start ​

  • Prepare the Kubernetes connection information, region, and availability zone.
  • Confirm the NPU vendor, model, memory, and device-plugin resource key.
  • Confirm whether the four cards are on one node or distributed across nodes.
  1. Create Regions and Availability Zones
  2. Maintain Accelerator Models
  3. Onboard the Cluster and Verify Devices
  4. Configure Metrics and Resource Specifications
  5. Build an Inference Template
  6. Configure Quotas and Monitoring

Document Index ​

DocumentDescription
Create Regions and Availability ZonesEstablish the shared resource boundary for clusters, specifications, images, and storage
Maintain Accelerator ModelsVerify the NPU model, memory, managed state, and Kubernetes resource key
Onboard the Cluster and Verify DevicesRegister the cluster and verify that all four NPU cards are reported
Configure Metrics and Resource SpecificationsCreate scheduling metrics and one-card, two-card, and four-card flavors
Regions & ZonesStart with the first operator page for On-Prem resource onboarding

Completion Checklist ​

Purpose: These are the scenario exit criteria. Use them to decide whether the outcome is observable and reviewable and whether you can continue to the next scenario. They do not repeat the procedure; if any item fails, return to the relevant feature guide and follow its troubleshooting section.

CheckPass Criteria
1The region and availability zone exist and are selectable during cluster registration.
2The cluster and nodes are available.
3AGIOne reports four target NPU cards.
4Inference templates or test workloads can select the new resource specifications.