Kiac v0.9.0 makes Apple GPU allocations easier to inspect, exposes DRA device metadata to workloads, and makes GPU startup and benchmark results more reliable.
kiac gpu inspect --name CLUSTERjoins DRA devices, reservations, share IDs, claims and consuming Pods. It explains pending workloads and supports-o json. Incomplete accounting is reported as unknown.- Containers requesting a DRA claim receive Kubernetes v1beta1 device metadata through a read-only mount. Applications can inspect the Venus device and their reservation without a Kubernetes API token. Try
examples/gpu-dra-metadata.yaml. kiac gpu benchnow requires confirmed model-layer offload and reportsoffloadedLayersseparately from the requested setting. Merely discovering a GPU no longer counts as a successful GPU benchmark.- GPU node bootstrap has bounded binary transfers and guest package processes. First-run package setup gets at least ten minutes, and bootstrap downloads use IPv4 to avoid the stalled mirror connection reproduced during hardware testing.
GPU memory reservations account for each VM's advertised window. They are not physical free-memory measurements or hard isolation, and multiple VMs share the Mac's physical GPU and unified memory.
Validated with full CI and real Apple M1 Max runs: K3s 1.37 DRA allocation/reallocation and stop/resume, K3s 1.36 DRA metadata, kubeadm 1.37 device-plugin workloads, and model-layer offload on Venus and native Metal. First-run Fedora downloads remain mirror-dependent. GPU status reports registration/device presence; it is not a continuous compute-health or memory-isolation guarantee.