inferenceGateway:
  enabled: true
  name: "inference-gateway"
  className: "istio"
  listeners:
    - name: http
      port: 80
      protocol: HTTP

global:
  image:
    pullPolicy: IfNotPresent
  imagePullSecrets: [] # 私有镜像仓库的 Secret，例如: [{"name": "registry-secret"}]

  env:
    - name: HF_HUB_OFFLINE # HuggingFace Hub 离线开关（1=离线；0=在线）
      value: "0"
    - name: PYTHONHASHSEED # cache-indexer 要求与所有目标 vLLM Pod 保持一致
      value: "0"
    - name: VLLM_KV_EVENTS_USE_INT_BLOCK_HASHES # 与 cache-indexer 的 BlockHash 归一化规则保持一致
      value: "1"

  modelName: "Qwen/Qwen3-8B" # 推理模型名称（用于 served-model-name 和标签）
  modelPath: "" # 推理模型本地路径（可选）。若配置，则 vLLM 使用该路径启动模型，否则使用 modelName。
  autoDownloadModel: true # 是否自动通过 huggingface (hf download) 下载模型，本地已有模型时建议关闭自动下载
  cachePath: "/home/llm_cache" # 推理侧缓存目录（如 HuggingFace / vLLM 的下载与编译缓存）

hermes-router: 
  enabled: true

  inferenceExtension:
    replicas: 1
    pluginsConfigFile: default-plugins.yaml
    routing:
      deploymentMode: aggregate
      profile: random
    tokenizer:
      enabled: false

  inferencePool:
    targetPorts:
      - number: 8000
    modelServerType: vllm

  httpRoute:
    inferenceGatewayName: "inference-gateway"

  provider:
    name: istio
    istio:
      destinationRule:
        trafficPolicy:
          tls:
            mode: SIMPLE         
            insecureSkipVerify: true
      retryConfig:
        enabled: true
        retryOn: "connect-failure,refused-stream,unavailable,cancelled,retriable-status-codes,5xx,reset"
        numRetries: 3

inference-backend:
  images: 
    inferenceEngine:
      repository: "hub.oepkgs.net/openfuyao/ascend/vllm-ascend"
      tag: "v0.18.0"

  inferenceDevice: "huawei.com/Ascend910"

  # 默认 volumeMounts（所有 vLLM deployment 共享）
  volumeMounts:
    ascend: # Ascend 设备相关的 volumeMounts
      enable: true

  # 默认 volumes（所有 vLLM deployment 共享）
  volumes:
    ascend: # Ascend 设备相关的 volumes
      enable: true

  services:
  - name: qwen3-8b-aggregated-0
    enabled: true
    mode: aggregated
    service:
      port: 8000
    kvTransferConfig: # KV Cache transfer 配置（aggregated 模式使用 store类型的 connector）
      connectorConfig:
        kv_connector: "AscendStoreConnector"
        kv_connector_extra_config:
          lookup_rpc_port: "1"
          backend: "mooncake"
      mooncake: # mooncake store提供的内存挂载配置
        configPath: "/app/mooncake.json"
        use_store: true
        master:
          rpcPort: 50051
          servicePort: 30089
          httpPort: 9003 # cache-indexer L3 依赖该 admin HTTP 端口（/get_all_keys、/query_key、/get_all_segments）
        # mooncake 配置文件内容（YAML 格式，initContainer 会自动替换 $POD_IP 并转换为 JSON）
        config: |
          local_hostname: "$POD_IP"
          metadata_server: "redis://redis-service:6379"
          master_server_address: "mooncake-master-service:30089"
          device_name: ""
          protocol: "ascend"
          global_segment_size: 42949672960
          use_ascend_direct: true

    aggregated:
      replicas: 2
      tensorParallelSize: 1
      pipelineParallelSize: 1
      dataParallelSize: 1
      dataParallelSizeLocal: 1
      dataParallelRpcPort: 12890
      env:
        - name: UB_SHARE_MEMORY
          value: "true"
      volumeMounts: [] # aggregated 节点特定的 volumeMounts
      volumes: [] # aggregated 节点特定的 volumes
      # extraArgs 是推理引擎 vLLM 自身的启动参数（模型配置与推理引擎调优参数），
      # 与部署拓扑无关；结构化字段仅保留 tensorParallelSize 等拓扑相关配置。
      extraArgs:
        - --max-model-len 10000
        - --max-num-batched-tokens 40960
        - --gpu-memory-utilization 0.8
        - --block-size 128
        - --trust-remote-code
        - --enable-prefix-caching
        - --prefix-caching-hash-algo sha256_cbor
        - --enable-prompt-tokens-details
        - --disable-access-log-for-endpoints=/health,/metrics
        # 默认过滤 vLLM 的 /health、/metrics 访问日志。
        # 该配置也会过滤对应接口的 4xx/5xx 请求日志，如需 Debug 请调整或移除该参数（--disable-access-log-for-endpoints）。

    resources: # aggregated 节点资源配置
      requests:
        cpu: "4"
        memory: "32Gi"
      limits:
        cpu: "8"
        memory: "64Gi"

# cache indexer
cache-indexer:
  enabled: false # 该 example 为 random example，不使用 cache-indexer
  resources: # cache-indexer 运行资源，生产环境可按集群规模调整
    requests:
      cpu: 100m
      memory: 128Mi
    limits:
      cpu: "1"
      memory: 2Gi
  blockKey: # 需与 vLLM 的 PYTHONHASHSEED / hash algo / int hash 开关保持一致
    pythonHashSeed: "0"
    prefixCachingHashAlgo: sha256_cbor
    useIntBlockHashes: true
  discovery:
    refreshInterval: 10s
    vllm:
      zmqPortName: zmq-pub
    mooncakeMaster:
      httpPortName: http
      httpPort: 9003
  service:
    name: cache-indexer-service
    port: 8080

eagle-eye:
  enabled: true

  network-performance-exporter:
    networkPerformanceExporter:
      metricsPort: 8222
      collectInterval: 15

pd-orchestrator:
  elastic-scaler:
    enabled: true # elastic-scaler 组件部署开关

    elasticScaler:
      enabled: true # chart包中默认提供的 elastic-scaler CR 实例开关，开启后会默认安装 elastic-scaler CR

    name: elastic-scaler # 通过 HPA 扩缩的最大副本数和最小副本数
    minReplicas: 1
    maxReplicas: 10

    targetRef:
      kind: resourcescalinggroup
      name: rsg # 该 name 需要对应 resourcescalinggroup.instanceConfig.name
      apiVersion: autoscaling.openfuyao.com/v1alpha1

    trigger:
      scalingAlgorithm: apa
      # 设置 apa 扩缩资源指标
      resource:
        metricsName: cpu
        targetType: Utilization
        targetValue: 70

  resourcescalinggroup:
    enabled: true # resourcescalinggroup 组件的部署开关

    instanceConfig:
      enabled: true # chart包中默认提供的 resource scaling group CR 实例开关，开启后会默认安装 CR 实例
      name: rsg
      targetResources:
      - name: aggregated
        resourceRef:
          apiVersion: leaderworkerset.x-k8s.io/v1
          kind: LeaderWorkerSet
          name: qwen3-8b-aggregated-0-aggregated # 该 name 需要和 inference backend 的 aggregated lws name 对应
      scalingStrategy:
        type: GroupReplication
        groupConfig: 
          groupName: aggregated 
          scaleDown: 
            metric: 
              expr: "vllm:num_requests_running" 
              type: "Gauge"
              window: "2m"
              aggregator: "avg"
            order: ascending

  tidal:
    enabled: true # tidal 组件的部署开关