Skip to content
lumenDocs

Inspect device resources

Available in v0.18.0. Open device resources in cluster navigation, or follow the device link from Pod or Node details. This is a read-only inspection workflow.

Prerequisites

Section titled “Prerequisites”

The cluster must expose resource.k8s.io/v1 (stable DRA, introduced in Kubernetes 1.34). Lumen does not fall back to beta APIs. A DRA driver must publish resources for inventory to appear; Lumen does not install drivers. Reads use the selected kubeconfig identity and Kubernetes RBAC.

Inspect a snapshot

Section titled “Inspect a snapshot”
  1. Select a context and namespace. The scope applies to claims, templates and Pods; classes and slices remain cluster-wide. Enter a known namespace if namespace discovery is denied.
  2. Select Claims, Templates, Classes, or Slices. Filter the list and open a resource to inspect its sanitized representation.
  3. Follow Pod claims to allocated devices, or inspect a Node’s published inventory and placement. Missing claims, pending allocations and unavailable inventory are different states.
  4. Use Refresh for a new observation. The capture time is not the driver’s health-update time; independent reads are not an atomic cluster snapshot.

Denied access, absent APIs, request failures and successful empty lists are shown separately. Readable sources remain useful when another source fails. Inventory can be incomplete because of pagination limits, missing slices, permissions or driver publication.

Lumen device-resource screen showing one synthetic published device slice

Actual main-build capture, 17 September 2026. This read-only workflow is available in v0.18.0. The device is synthetic; no GPU driver or physical hardware is involved.

What the data means

Section titled “What the data means”

Health comes from matching Kubernetes per-container reports. Missing reports are unknown; terminated Pods can retain old reports. Lumen does not probe hardware or measure GPU utilization. A published device is not necessarily unused: sharing, administrative allocation and limited namespace visibility prevent a reliable global free-device count.

The inspector removes annotations and managed fields and redacts opaque driver data. Its JSON is valid YAML syntax, but is not a complete manifest export for replay. Synthetic cluster tests establish API and RBAC behavior, not physical hardware allocation or real driver health.

Sources: device resource guide, src/routes/cluster/DeviceResourcesView.tsx, src/lib/deviceResources.ts, and src-tauri/src/k8s/devices.rs at v0.18.0.