servingEngineSpec:
  # Global runtime class for all models (can be overridden per model)
  runtimeClassName: "nvidia"

  # Global tolerations applied to all model deployments
  tolerations:
    - key: "gpu-node"
      operator: "Exists"
      effect: "NoSchedule"

  modelSpec:
  - name: "opt125m"
    repository: "lmcache/vllm-openai"
    tag: "latest"
    modelURL: "facebook/opt-125m"
    # Override global runtimeClassName for this specific model
    runtimeClassName: "custom-runtime"

    replicaCount: 1

    requestCPU: 6
    requestMemory: "16Gi"
    requestGPU: 1

    pvcStorage: "10Gi"

    # Per-model tolerations: appended to the global tolerations (default policy)
    tolerations:
      - key: "dedicated"
        operator: "Equal"
        value: "opt125m"
        effect: "NoSchedule"
    # tolerationsPolicy: "append"  # default: union of global + model tolerations
    # tolerationsPolicy: "override" # replaces global tolerations entirely for this model
cacheserverSpec:
  repository: "lmcache/vllm-openai"
  tag: "latest-nightly"
  containerPort: 8000
  servicePort: 80
  serde: "naive"
  resources:
    requests:
      cpu: "4"
      memory: "8Gi"
      # Example InfiniBand/RDMA extended resource
      rdma/ib: "1"
    limits:
      cpu: "4"
      memory: "10Gi"
      # Example InfiniBand/RDMA extended resource
      rdma/ib: "1"