kiac v0.7.1 - Real Apple GPU clusters (alpha)

Real Apple GPU clusters (alpha)

Kiac can now create multi-node Kubernetes clusters with dedicated workers backed by the real Apple GPU through krunkit, virtio-gpu, Mesa Venus, and Vulkan. Ordinary clusters remain on the fast apple/container backend and pay no GPU-mode startup or idle overhead.

Highlights

• Create GPU clusters with either K3s or kubeadm using --gpu-workers N.
• Schedule the honest kiac.dev/gpu resource through the default device plugin or Kubernetes DRA on Kubernetes 1.36+.
• Use DRA ResourceClaims for consumable GPU-memory accounting and CDI-based /dev/dri injection.
• Inspect and prove the path with kiac gpu doctor, status, bench, values, and namespace-scoped compat commands.
• Keep the normal Kiac experience: Cilium or kindnet, LoadBalancer Services, Gateway API, storage, metrics, observability, image loading, stop/start, resume, verify, support bundles, and multi-node traffic.
• Follow the complete GPU and inference lab, including Vulkan proof, DRA claims, TinyLlama inference, and native Metal-versus-Venus benchmarking.

Important boundaries

• GPU mode is alpha and currently supports Apple silicon Macs with IPv4 clusters.
• Linux pods receive the real Apple GPU through Venus/Vulkan. They do not receive CUDA, NVML, nvidia-smi, Metal, or MLX.
• Kiac does not advertise nvidia.com/gpu capacity. The optional compatibility webhook only rewrites legacy resource requests inside explicitly enabled namespaces.
• GPU clusters require krunkit 1.3.2+ and vmnet-helper 0.13.0+; kiac gpu doctor checks the complete host toolchain.

Upgrade

brew update
brew upgrade kiac
kiac gpu doctor

Full changelog: v0.6.0...v0.7.1