deployment.yaml

apiVersion: apps/v1
kind: Deployment
metadata:
  name: {{ include "server.llama.cpp.fullname" . }}
  labels:
    {{- include "server.llama.cpp.labels" . | nindent 4 }}
spec:
  {{- if not .Values.autoscaling.enabled }}
  replicas: {{ .Values.replicaCount }}
  {{- end }}
  selector:
    matchLabels:
      {{- include "server.llama.cpp.selectorLabels" . | nindent 6 }}
  template:
    metadata:
      annotations:
        {{- include "server.llama.cpp.labels" . | nindent 8 }}
        {{- if .Values.server.metrics }}
        prometheus.io/scrape: 'true'
        prometheus.io/port: '{{ .Values.server.port }}'
        {{- end }}
      {{- with .Values.podAnnotations }}
        {{- toYaml . | nindent 8 }}
      {{- end }}
      labels:
        prometheus.io/scrape: 'true'
        {{- include "server.llama.cpp.labels" . | nindent 8 }}
        {{- with .Values.podLabels }}
        {{- toYaml . | nindent 8 }}
        {{- end }}
    spec:
      {{- with .Values.imagePullSecrets }}
      imagePullSecrets:
        {{- toYaml . | nindent 8 }}
      {{- end }}
      {{- with .Values.nodeSelector }}
      nodeSelector:
        {{- toYaml . | nindent 8 }}
      {{- end }}
      {{- with .Values.affinity }}
      affinity:
        {{- toYaml . | nindent 8 }}
      {{- end }}
      {{- with .Values.tolerations }}
      tolerations:
        {{- toYaml . | nindent 8 }}
      {{- end }}
      containers:
        - name: {{ .Chart.Name }}
          securityContext:
            {{- toYaml .Values.securityContext | nindent 12 }}
          image: "{{ .Values.images.server.repository }}:{{ .Values.images.server.name }}-{{ .Values.images.server.tag | default .Chart.AppVersion }}"
          imagePullPolicy: {{ .Values.images.pullPolicy }}
          resources:
            {{- toYaml .Values.resources | nindent 12 }}
          command:
            - {{ .Values.server.command }}
          args:
            - --host
            - {{ .Values.server.host }}
            - --port
            - "{{ .Values.server.port }}"
            - --model
            - {{ .Values.model.path }}/{{  regexReplaceAll "(.*/)?([^/]+).gguf" .Values.model.file "${2}.gguf" }}
            - --cont-batching
            - --alias
            - {{ .Values.model.alias }}
            - --ctx-size
            - "{{ .Values.server.kvCache.size }}"
            - --parallel
            - "{{ .Values.server.slots }}"
            {{- if .Values.server.embeddings }}
            - --embedding
            {{- end }}
            {{- if .Values.server.metrics }}
            - --metrics
            {{- end }}
            - --log-format
            - {{ .Values.server.log.format }}
            {{- if  .Values.server.log.disabled }}
            - --log-disable
            {{- end }}
            {{- with .Values.server.extraArgs }}
            {{- toYaml . | nindent 12 }}
            {{- end }}
          ports:
            - name: http
              containerPort: {{ .Values.server.port }}
              protocol: TCP
          {{- with .Values.volumeMounts }}
          volumeMounts:
            {{- toYaml . | nindent 12 }}
          {{- end }}
          volumeMounts:
            - mountPath: {{ .Values.model.path }}
              name: models
              readOnly: true
      volumes:
        - name: models
          persistentVolumeClaim:
            claimName: {{ include "server.llama.cpp.fullname" . }}
            readOnly: true