# Plain vLLM Deployment (no KServe) — real-GPU serving for the Phase 2 numbers.
# Serves an OpenAI-compatible API + Prometheus vllm:* metrics on :8000.
# Pinned to the laptop 4070 (most free VRAM) for the single-replica baseline.
# Model is parameterised via the MODEL arg block so the same manifest does 1.5B then 3B.
apiVersion: v1
kind: Namespace
metadata:
  name: inference
---
apiVersion: apps/v1
kind: Deployment
metadata:
  name: vllm-qwen
  namespace: inference
  labels:
    app: vllm-qwen
spec:
  replicas: 1
  strategy:
    type: Recreate          # single GPU per node — can't run old+new pod at once
  selector:
    matchLabels:
      app: vllm-qwen
  template:
    metadata:
      labels:
        app: vllm-qwen
    spec:
      runtimeClassName: nvidia
      nodeSelector:
        kubernetes.io/hostname: kaiser-laptop      # 4070, ~7.9 GB free
      # hostNetwork + Default DNS: the cross-node pod overlay is down under WSL2
      # mirrored mode, so pods can't reach CoreDNS (which lives on the other node)
      # and can't resolve huggingface.co. Using the host's DNS sidesteps that and
      # also makes :8000 scrapeable by node IP. See docs/cluster-troubleshooting-log.md
      hostNetwork: true
      dnsPolicy: Default
      containers:
        - name: vllm
          image: vllm/vllm-openai:latest
          args:
            - --model=Qwen/Qwen2.5-1.5B-Instruct
            - --served-model-name=qwen
            - --gpu-memory-utilization=0.85          # fit within free VRAM (not total)
            - --max-model-len=8192                    # cap KV-cache footprint
            - --port=8000
          ports:
            - name: http
              containerPort: 8000
          resources:
            limits:
              nvidia.com/gpu: 1
          volumeMounts:
            - name: dshm
              mountPath: /dev/shm
          readinessProbe:
            httpGet:
              path: /health
              port: 8000
            initialDelaySeconds: 40
            periodSeconds: 10
            failureThreshold: 90                      # allow time for image pull + model download
      volumes:
        - name: dshm
          emptyDir:
            medium: Memory
            sizeLimit: 2Gi
---
apiVersion: v1
kind: Service
metadata:
  name: vllm-qwen
  namespace: inference
  labels:
    app: vllm-qwen
spec:
  selector:
    app: vllm-qwen
  ports:
    - name: http
      port: 8000
      targetPort: 8000
---
apiVersion: monitoring.coreos.com/v1
kind: PodMonitor
metadata:
  name: vllm-qwen
  namespace: inference
  labels:
    release: kube-prometheus-stack        # match the Prometheus podMonitorSelector
spec:
  selector:
    matchLabels:
      app: vllm-qwen
  podMetricsEndpoints:
    - port: http
      path: /metrics
      interval: 15s