Qwen/Qwen2.5-0.5B on CPU

# An InferenceClass describing a Civo CPU node pool - no GPU at all.
#
# The provisioning block has no accelerator, so the cluster composition
# builds the pool with the CPU role: a standard node image and the
# modelplane.ai/cpu label and taint, nothing GPU. The serving stack installs
# dra-driver-cpu when a class claims dra.cpu, which publishes each node's
# cores as claimable devices (one device per NUMA node) and registers the
# dra.cpu DeviceClass.
apiVersion: modelplane.ai/v1alpha1
kind: InferenceClass
metadata:
  name: civo-cpu-8vcpu
spec:
  description: "Civo g4p.kube.medium, 8 vCPUs, no GPU"
  provisioning:
    provider: Civo
    civo:
      size: g4p.kube.medium
  devices:
  - name: cpu
    claim: DRA
    driver: dra.cpu
    deviceClassName: dra.cpu
    count: 1
    capacity:
      # The node's 8 vCPUs less the two dra-driver-cpu reserves for the OS
      # and kubelet (reservedCPUs 0-1), which is what the driver's
      # ResourceSlices report as claimable.
      cpu: { value: "6" }

# A Civo cluster in NYC1 with a single CPU node offering the class above.
# The composed ServingStack gets draDrivers: [dra.cpu], so dra-driver-cpu
# installs and the GPU device stacks stay out - a cluster with no GPU runs
# no GPU machinery.
apiVersion: modelplane.ai/v1alpha1
kind: InferenceCluster
metadata:
  name: civo-cpu
  labels:
    modelplane.ai/region: nyc1
spec:
  cluster:
    source: Civo
    civo:
      region: NYC1
      # CPU cores are claimed as consumable capacity, which Kubernetes
      # enables by default from 1.36 (DRAConsumableCapacity) - on an older
      # version the engine pod fails at NodePrepareResources. Civo's
      # default version may be older, so a CPU cluster pins it; list the
      # exact strings with: civo kubernetes versions.
      kubernetesVersion: 1.36.0-k3s1
  nodePools:
  - name: cpu-pool
    className: civo-cpu-8vcpu
    nodeCount: 1

# A small model served on CPUs alone. The device selector matches against
# the claimable cores the civo-cpu-8vcpu class declares, so >= 4 lands it on
# the CPU pool; the claim takes the engine's CPU request, five cores, as
# capacity from the NUMA-node device and pins the engine container to a
# cpuset. The engine is vLLM's CPU build - the default
# vllm/vllm-openai image is CUDA-only. --block-size=32 is the smallest
# block its CPU attention backend accepts (multiples of 32); Modelplane
# reads the flag and chunks the endpoint picker's prefix cache the same
# way, so routing stays aligned. VLLM_CPU_KVCACHE_SPACE sets the KV
# cache's share of the node's RAM in GiB. No ModelCache: the engine pulls
# weights straight from Hugging Face.
apiVersion: modelplane.ai/v1alpha1
kind: ModelDeployment
metadata:
  name: qwen-cpu-demo
  namespace: ml-team
spec:
  replicas: 1
  template:
    spec:
      engines:
      - name: qwen
        members:
        - role: Standalone
          nodeSelector:
            devices:
            - name: cpu
              count: 1
              selectors:
              - cel: |
                  device.capacity["dra.cpu"].cpu.compareTo(quantity("4")) >= 0
          template:
            spec:
              containers:
              - name: engine
                image: vllm/vllm-openai-cpu:v0.32.0-x86_64
                args:
                - --model=Qwen/Qwen2.5-0.5B-Instruct
                - --served-model-name=$(MODELPLANE_SERVED_MODEL_NAME)
                - --max-model-len=4096
                - --block-size=32
                env:
                - name: VLLM_CPU_KVCACHE_SPACE
                  value: "4"
                # Sizes the pod's accounting and the DRA claim alike:
                # compose-model-replica stamps this as the claim's capacity
                # request, and the kubelet accounts CPUs only from requests
                # until KEP-5517. It must stay below the device's six
                # claimable cores - a whole-device claim exhausts the
                # shared CPU pool and the driver refuses to prepare it
                # while the node's DaemonSets need somewhere to run. Five
                # leaves them one claimable core.
                resources:
                  requests:
                    cpu: "5"
                  limits:
                    cpu: "5"

# Publishes the deployment's endpoints as one model a caller can name:
# <namespace>/<service>, so ml-team/qwen-cpu here. Modelplane composes one
# ModelEndpoint per ModelReplica, labeled with the deployment name, so
# this selector reaches every replica.
apiVersion: modelplane.ai/v1alpha1
kind: ModelService
metadata:
  name: qwen-cpu
  namespace: ml-team
spec:
  endpoints:
  - name: qwen-cpu-demo
    selector:
      matchLabels:
        modelplane.ai/deployment: qwen-cpu-demo