vllm/.helm/templates/sts.yaml

---
apiVersion: apps/v1
kind: StatefulSet
metadata:
  name: {{ .Chart.Name }}
  labels:
    app: {{ .Chart.Name }}
spec:
  serviceName: {{ .Chart.Name }}
  minReadySeconds: 10
  updateStrategy:
    type: RollingUpdate
  replicas: 1
  selector:
    matchLabels:
      app: {{ .Chart.Name }}
  template:
    metadata:
      labels:
        app: {{ .Chart.Name }}
    spec:
      affinity:
        podAntiAffinity:
          preferredDuringSchedulingIgnoredDuringExecution:
          - podAffinityTerm:
              labelSelector:
                matchLabels:
                  app: {{ .Chart.Name }}
              topologyKey: kubernetes.io/hostname
            weight: 100
      {{- if and .Values.gpu.enabled .Values.gpu.nodeSelector }}
      nodeSelector: {{ toYaml .Values.gpu.nodeSelector | nindent 8 }}
      {{- end }}
      {{- if and .Values.gpu.enabled .Values.gpu.tolerations }}
      tolerations: {{ toYaml .Values.gpu.tolerations | nindent 6 }}
      {{- end }}
      imagePullSecrets:
      - name: registrysecret
      containers:
        - name: {{ .Chart.Name }}
          image: {{ .Values.werf.image.vllm }}
          imagePullPolicy: IfNotPresent
          volumeMounts:
            - name: data
              mountPath: /root/.cache/huggingface
          ports:
            - name: http
              containerPort: {{ .Values.port }}
              protocol: TCP
          args:
          - {{ .Values.env.VLLM_MODEL }}
          - --async-scheduling
          - --host
          - 0.0.0.0
          - --port
          - "{{ .Values.port }}"
          - --gpu-memory-utilization
          - "0.2"
          - --max-model-len
          - "{{ .Values.env.VLLM_MAX_MODEL_LEN }}"
          - --enforce-eager
          - --tensor-parallel-size
          - "1"
          - --trust-remote-code
          - --chat-template-content-format
          - "string"
          env:
          {{- range $n, $v := .Values.env }}
            - name: {{ $n }}
              value: {{ $v | quote }}
          {{- end }}
          livenessProbe:
            httpGet:
              path: /health
              port: {{ .Values.port }}
            initialDelaySeconds: 120
            periodSeconds: 9
            timeoutSeconds: 6
            failureThreshold: 3
          readinessProbe:
            httpGet:
              path: /v1/models
              port: {{ .Values.port }}
            initialDelaySeconds: 120
            periodSeconds: 5
            timeoutSeconds: 6
            failureThreshold: 3
          resources: {{ dig .Values.werf.env .Values.resources._default .Values.resources | toYaml | nindent 12 }}
  volumeClaimTemplates:
  - metadata:
      name: data
    spec:
      accessModes: [ "ReadWriteOnce" ]
      storageClassName: {{ dig .Values.werf.env .Values.storage.class._default .Values.storage.class | quote }}
      resources:
        requests:
          storage: {{ dig .Values.werf.env .Values.storage.size._default .Values.storage.size | quote }}
---
apiVersion: v1
kind: Service
metadata:
  name: {{ .Chart.Name }}
  labels:
    app: {{ .Chart.Name }}
    prometheus.deckhouse.io/custom-target: {{ .Chart.Name }}
  annotations:
    prometheus.deckhouse.io/port: "{{ .Values.port }}"
    prometheus.deckhouse.io/path: "/metrics/"
spec:
  type: ClusterIP
  ports:
    - port: {{ .Values.port }}
      targetPort: http
      protocol: TCP
      name: http
  selector:
    app: {{ .Chart.Name }}