# Three vLLM replicas, one GPU each, one per node. The label llm-d.ai/role=aggregated is what the
# InferencePool (30-inferencepool.yaml) and rule.sh select on.
apiVersion: apps/v1
kind: Deployment
metadata:
  name: pd-disaggregation-nvidia-gpu-vllm-aggregated
  namespace: llm-d-pd-disaggregation
  labels:
    llm-d.ai/guide: pd-disaggregation
    llm-d.ai/role: aggregated
spec:
  replicas: 3
  selector:
    matchLabels:
      llm-d.ai/role: aggregated
  template:
    metadata:
      labels:
        llm-d.ai/engine-type: vllm
        llm-d.ai/guide: pd-disaggregation
        llm-d.ai/inference-serving: "true"
        llm-d.ai/model: Qwen3-32B-FP8
        llm-d.ai/role: aggregated
    spec:
      affinity:
        podAntiAffinity:
          requiredDuringSchedulingIgnoredDuringExecution:
          - labelSelector:
              matchLabels:
                llm-d.ai/role: aggregated
            topologyKey: kubernetes.io/hostname
      containers:
      - name: modelserver
        image: docker.io/vllm/vllm-openai:v0.30.0
        imagePullPolicy: IfNotPresent
        command: ["vllm", "serve"]
        args:
        - Qwen/Qwen3-32B-FP8
        - --port=8000
        - --disable-access-log-for-endpoints=/health,/metrics,/v1/models
        - --tensor-parallel-size=1
        - --block-size=16                           # the EPP's prefix-cache model uses 16-token blocks
        - --prefix-caching-hash-algo=sha256_cbor
        - --enable-prompt-tokens-details           # usage.prompt_tokens_details.cached_tokens in every answer: the prefix-cache hit of that request
        - --kv-events-config
        - '{"enable_kv_cache_events":true,"publisher":"zmq","endpoint":"tcp://*:5557"}'
        env:
        - name: HF_TOKEN
          valueFrom:
            secretKeyRef:
              name: llm-d-hf-token
              key: HF_TOKEN
              optional: true                        # Qwen3-32B-FP8 is public; set the secret for gated models
        - name: PYTHONHASHSEED
          value: "0"
        - name: VLLM_KV_EVENTS_USE_INT_BLOCK_HASHES
          value: "1"
        - name: VLLM_HTTP_TIMEOUT_KEEP_ALIVE
          value: "120"
        ports:
        - name: modelserver
          containerPort: 8000
          protocol: TCP
        - name: kv-events
          containerPort: 5557
          protocol: TCP
        startupProbe:                               # weight download + compile on first start
          httpGet:
            path: /v1/models
            port: modelserver
          initialDelaySeconds: 15
          periodSeconds: 30
          timeoutSeconds: 5
          failureThreshold: 120
        readinessProbe:
          httpGet:
            path: /v1/models
            port: modelserver
          periodSeconds: 5
          timeoutSeconds: 2
          failureThreshold: 3
        livenessProbe:
          httpGet:
            path: /health
            port: modelserver
          periodSeconds: 10
          timeoutSeconds: 5
          failureThreshold: 3
        resources:
          requests:
            cpu: "4"
            memory: 32Gi
            nvidia.com/gpu: "1"
          limits:
            cpu: "8"
            memory: 64Gi
            nvidia.com/gpu: "1"
        volumeMounts:
        - name: shm
          mountPath: /dev/shm
        - name: torch-compile-cache
          mountPath: /.cache
      volumes:
      - name: shm
        emptyDir:
          medium: Memory
          sizeLimit: 16Gi
      - name: torch-compile-cache
        emptyDir: {}