Device Plugin and DRA allocation paths

Device Plugin and DRA allocation paths An architecture diagram generated by Archify. Pod workload · device request / execution · Architecture component Pod workload device request / execution ResourceClaim · DeviceClass / exactly request · Architecture component ResourceClaim DeviceClass / exactly request Kubernetes scheduler · claim / Pod placement constraints · Architecture component Kubernetes scheduler claim / Pod placement constraints Device Plugin · nvidia.com/gpu / MIG resources · Architecture component Device Plugin nvidia.com/gpu / MIG resources ResourceSlice · driver-published inventory · Architecture component ResourceSlice driver-published inventory NVIDIA DRA driver · prepare allocated devices / CDI · Architecture component NVIDIA DRA driver prepare allocated devices / CDI reference evaluate requests attributes / capacity prepare allocation publish inventory extended-resource path Legend Backend Database External

Sharing and feature maturity

  • • Device plugins also provide configured MIG/time-slicing paths.
  • • DRA API maturity differs from MPS/DynamicMIG/TimeSlicing gates.
  • • Handwritten ResourceSlices do not create GPUs.

Deployment paths

  • • Operator26.7:GPUCluster and ClusterPolicy are mutually exclusive.
  • • Standalone0.5 rejects default GPU rendering until explicit opt-in.
  • • Avoid competing device-plugin/DRA ownership of the same devices.