apiVersion: v1
kind: PersistentVolumeClaim
metadata:
  name: vllm-models
spec:
  accessModes:
    - ReadWriteOnce
  volumeMode: Filesystem
  resources:
    requests:
      storage: 50Gi
---
apiVersion: apps/v1
kind: Deployment
metadata:
  name: vllm-server
spec:
  replicas: 1
  selector:
    matchLabels:
      app.kubernetes.io/name: vllm
  template:
    metadata:
      labels:
        app.kubernetes.io/name: vllm
        app: vllm
        workload-type: inference
      annotations:
        prometheus.io/scrape: 'true'
        prometheus.io/port: '8001'
        prometheus.io/path: '/metrics'
    spec:
      # Removed nodeSelector for GPU nodes as they don't appear to exist in the cluster
      # If you have GPU nodes with a different label, you can uncomment and modify this section
      # nodeSelector:
      #   <your-gpu-node-label-key>: <your-gpu-node-label-value>
      containers:
      - name: vllm
        image: vllm/vllm-openai:latest
        command: ["/bin/sh", "-c"]
        args:
        - "vllm serve ${INFERENCE_MODEL} --enforce-eager --max-model-len 8192 --gpu-memory-utilization 0.7 --enable-auto-tool-choice --tool-call-parser llama3_json --max-num-seqs 4 --port 8001"
        env:
        - name: INFERENCE_MODEL
          value: "${INFERENCE_MODEL}"
        - name: HUGGING_FACE_HUB_TOKEN
          valueFrom:
            secretKeyRef:
              name: hf-token-secret
              key: token
        ports:
          - containerPort: 8001
            name: http
        resources:
          limits:
            nvidia.com/gpu: 1
          requests:
            nvidia.com/gpu: 1
        volumeMounts:
          - name: llama-storage
            mountPath: /root/.cache/huggingface
      volumes:
      - name: llama-storage
        persistentVolumeClaim:
          claimName: vllm-models
---
apiVersion: v1
kind: Service
metadata:
  name: vllm-server
spec:
  selector:
    app.kubernetes.io/name: vllm
  ports:
  - protocol: TCP
    port: 8001
    targetPort: 8001
    name: http
  type: ClusterIP