apiVersion: catalyst.openova.io/v1alpha1
kind: Blueprint
metadata:
  name: bp-vllm
  labels:
    catalyst.openova.io/category: ai-runtime
    catalyst.openova.io/section: pts-4-6-llm-serving
spec:
  version: 1.0.0
  card:
    title: vLLM
    summary: High-throughput LLM inference engine with PagedAttention. OpenAI-compatible API. GPU-accelerated when nvidia.com/gpu is available; CPU fallback for non-GPU dev Sovereigns.
    icon: vllm.svg
    category: ai-runtime
    tags: [llm, inference, openai-compatible, gpu, ai]
    documentation: https://docs.vllm.ai/
    license: Apache-2.0
  visibility: listed
  owner:
    team: ai-platform
    contact: ai-platform@openova.io
  configSchema:
    type: object
    properties:
      model:
        type: string
        default: "meta-llama/Llama-3.1-8B-Instruct"
        description: HuggingFace model ID or in-cluster path served by vLLM.
      replicas:
        type: integer
        default: 1
        minimum: 1
        maximum: 16
      gpu:
        type: object
        properties:
          enabled:
            type: boolean
            default: false
            description: Set true on a GPU-equipped Sovereign. When false, vLLM runs on CPU (dev only — not for production traffic).
          count:
            type: integer
            default: 1
            description: Number of `nvidia.com/gpu` units to request when gpu.enabled=true.
      maxModelLen:
        type: integer
        default: 8192
        description: Maximum context length passed to vLLM via --max-model-len.
      gpuMemoryUtilization:
        type: number
        default: 0.9
        description: Fraction of GPU memory vLLM may use (--gpu-memory-utilization).
  placementSchema:
    modes: [single-region, active-active]
    default: single-region
  manifests:
    chart: ./chart
  depends:
    - blueprint: bp-kserve
      version: ^1.0
      alias: kserve
  upgrades:
    from: ["0.x"]
  observability:
    metrics: prometheus
    logs: stdout