05-model-release.yaml

apiVersion: ml.k8s-lab.io/v1alpha1
kind: ModelRelease
metadata:
  name: qwen2.5-7b-k8s-chat-prod-20260413
spec:
  owner: platform-llm
  stage: canary
  modelVersionRef:
    name: qwen2.5-7b-k8s-sft-lora-v1
  runtimePackage:
    uri: "s3://ml-assets/runtime/qwen2.5-7b-k8s-sft-v1-awq"
    digest: "sha256:3d2ef71c8b6f0a0b8c-example"
  engineConfigRef:
    path: "ml-platform/examples/21-llm-serving/00-inference-service-config.yaml"
  deploymentRef:
    name: "qwen2.5-7b-k8s-chat"
    namespace: "llm-serving"
  approvals:
    offlineEval: approved
    loadTest: approved
    securityScan: approved
    productOwner: approved
  rollback:
    targetRelease: "qwen2.5-7b-k8s-chat-prod-20260406"
    maxRollbackTimeMinutes: 10
  trafficPolicy:
    initialCanaryPercent: 10
    promoteIf:
      - "ttft_p95_ms < 1800"
      - "error_rate < 0.02"
      - "user_abort_rate delta < 0.03"

notes:
  - "这是服务版本发布单，不是训练阶段的模型登记卡。"
  - "这里同时引用 modelVersion、runtimePackage 和 deployment。"