apiVersion: catalog.confighub.com/v1alpha1
kind: SourceAndIntentRecord
metadata:
  name: eks-inference-vllm-cpu-starter
spec:
  purpose: Run one real OpenAI-compatible model request without requiring a GPU or cloud account.
  source:
    type: existing-oci
    reference: oci://ghcr.io/confighub/configs/eks-inference/inference-workloads
    digest: sha256:4ae5737725c86f7a07cf3366a7eafee8625a0dd8487784aa581ce7e33d460eb2
    producerCommit: f6a1689c2fe231d1bc2262d9916aa69a774882b4
    receipt: data/certified-bundles/receipts/eks-inference/inference-workloads/receipt.yaml
  selection:
    includedUnits:
      - 00-namespace
      - chat
      - vllm-qwen
    retainedInactiveUnits:
      - smoke-cpu
      - smoke-gpu
  changes:
    - unit: vllm-qwen
      reason: Replace the L4 and 7B AWQ runtime with a small CPU profile for an accessible functional test.
      fields:
        - spec.replicas
        - spec.template.spec.nodeSelector
        - spec.template.spec.tolerations
        - spec.template.spec.containers[name=vllm].image
        - spec.template.spec.containers[name=vllm].args
        - spec.template.spec.containers[name=vllm].resources
    - unit: chat
      reason: Send one deterministic request to the model selected by the CPU profile.
      fields:
        - data.request
        - spec.template.spec.containers[name=chat].env
  runtime:
    image:
      reference: docker.io/vllm/vllm-openai-cpu
      tag: v0.27.1-arm64
      digest: sha256:e6745d7ba6610f637c6f22fc06cd730342e50245b6c46767235600483adfbbde
      architecture: arm64
      device: cpu
    model:
      identity: Qwen/Qwen2.5-0.5B-Instruct
      revision: 7ae557604adf67be50417f59c2c2f167def9a775
      access: public
      credentialsRequired: false
    request:
      api: OpenAI-compatible chat completions
      prompt: What is two plus two? Reply with only the number.
      success: The response names the selected model, has a non-empty answer, and finishes without an API error.
    targetSettings:
      memoryLimit: 10Gi
      memoryReservationFraction: 0.2
      reason: Keep vLLM's CPU cache reservation below the pod limit on a 32 GiB kind node.
  lifecycle:
    order:
      - Create the inference namespace.
      - Start vLLM and wait for its health endpoint.
      - Start the in-cluster client.
      - Run and record the model request.
    retry: A failed model download or startup leaves the Deployment retryable; a request is attempted only after rollout readiness.
    cleanup: Scale or delete the vLLM Deployment when the test is complete.
  boundaries:
    proves:
      - A pinned vLLM CPU image starts on an ARM64 target.
      - A pinned public model revision loads.
      - One OpenAI-compatible model request returns a non-empty answer.
    doesNotProve:
      - NVIDIA GPU readiness
      - Qwen 7B AWQ performance or correctness
      - EKS or Karpenter provisioning
      - production capacity, latency, or response quality