spec.capacity.hami.mode: DRA is accepted but the provider returns ErrUnsupported from Devices and PendingDemand, so nothing drives off it. Only DevicePlugin mode is implemented.
DRA is the cleaner accounting model — capacity is a first-class API field (ResourceSlice.devices[].capacity, ResourceClaim.status.allocation.devices[].consumedCapacity) rather than an annotation protocol — and docs/design/gpucellpool-capacity.md §3 already specifies the mapping.
Blocked on an environment, not on design: needs a workload cluster on Kubernetes >= 1.34 with the DRAConsumableCapacity feature gate, CDI, and projecthami/k8s-dra-driver. The provider should preflight those and report CapacityProviderReady=False/DRAFeatureGateMissing rather than reading empty slices, because a missing gate looks exactly like "no capacity".
spec.capacity.hami.mode: DRAis accepted but the provider returnsErrUnsupportedfromDevicesandPendingDemand, so nothing drives off it. Only DevicePlugin mode is implemented.DRA is the cleaner accounting model — capacity is a first-class API field (
ResourceSlice.devices[].capacity,ResourceClaim.status.allocation.devices[].consumedCapacity) rather than an annotation protocol — anddocs/design/gpucellpool-capacity.md§3 already specifies the mapping.Blocked on an environment, not on design: needs a workload cluster on Kubernetes >= 1.34 with the
DRAConsumableCapacityfeature gate, CDI, andprojecthami/k8s-dra-driver. The provider should preflight those and reportCapacityProviderReady=False/DRAFeatureGateMissingrather than reading empty slices, because a missing gate looks exactly like "no capacity".