# services/hermes/local-image-deployment.yaml
apiVersion: apps/v1
kind: Deployment
metadata:
  name: hermes-local-image
  namespace: hermes
  labels:
    app: hermes-local-image
spec:
  replicas: 1
  revisionHistoryLimit: 2
  strategy:
    type: Recreate
  selector:
    matchLabels:
      app: hermes-local-image
  template:
    metadata:
      labels:
        app: hermes-local-image
      annotations:
        ai.bstein.dev/model: black-forest-labs/FLUX.2-klein-4B
        ai.bstein.dev/gpu: titan-24 lease-shared image, desktop, and Wolf lane
        ai.bstein.dev/config-rev: "20260811-flux2-klein-low-vram"
    spec:
      serviceAccountName: hermes-gpu-runtime
      runtimeClassName: nvidia
      affinity:
        nodeAffinity:
          requiredDuringSchedulingIgnoredDuringExecution:
            nodeSelectorTerms:
              - matchExpressions:
                  - key: kubernetes.io/hostname
                    operator: In
                    values:
                      - titan-24
      volumes:
        - name: models
          persistentVolumeClaim:
            claimName: hermes-image-models
        - name: local-image-tmp
          emptyDir:
            sizeLimit: 2Gi
      containers:
        - name: local-image
          image: registry.bstein.dev/bstein/hermes-local-image@sha256:d6257f49a60244fb1e30733848a187fa932acebfdb82a2560d4e760fa018b645
          imagePullPolicy: IfNotPresent
          ports:
            - name: local-image
              containerPort: 9004
          env:
            - name: HERMES_LOCAL_IMAGE_MODEL
              value: black-forest-labs/FLUX.2-klein-4B
            - name: HERMES_LOCAL_IMAGE_REVISION
              value: e7b7dc27f91deacad38e78976d1f2b499d76a294
            - name: HERMES_LOCAL_IMAGE_LISTEN_PORT
              value: "9004"
            - name: LEASE_NAMESPACE
              value: hermes
            - name: LEASE_NAME
              value: titan-24-gpu-owner
            - name: LEASE_IDLE_OWNER
              value: hermes
            - name: LEASE_IMAGE_OWNER
              value: hermes-image
            - name: HERMES_LOCAL_IMAGE_GPU_ACTIVITY_URL
              value: http://nvidia-process-exporter-local.monitoring.svc.cluster.local:9401/metrics
            - name: HERMES_LOCAL_IMAGE_GPU_ACTIVITY_NODE
              value: titan-24
            # The lease is authoritative for checkout state. Keep a high SM
            # ceiling only as a fail-safe for unregistered GPU workloads;
            # desktop compositors can briefly spike well above their idle use.
            - name: HERMES_LOCAL_IMAGE_GPU_ACTIVE_SM_PERCENT
              value: "80"
            - name: HERMES_LOCAL_IMAGE_GPU_MAX_EXTERNAL_MEMORY_BYTES
              value: "3221225472"
            - name: HERMES_LOCAL_IMAGE_GPU_ACTIVITY_SAMPLES
              value: "3"
            - name: HERMES_LOCAL_IMAGE_GPU_ACTIVITY_SAMPLE_INTERVAL
              value: "1"
            - name: HERMES_LOCAL_IMAGE_RENDER_TIMEOUT
              value: "1200"
            - name: HERMES_LOCAL_IMAGE_OFFLOAD_MODE
              value: sequential
            - name: PYTORCH_CUDA_ALLOC_CONF
              value: expandable_segments:True
            - name: HF_HOME
              value: /models/huggingface
            - name: HOME
              value: /tmp
            - name: NVIDIA_VISIBLE_DEVICES
              value: all
            - name: NVIDIA_DRIVER_CAPABILITIES
              value: compute,utility
          volumeMounts:
            - name: models
              mountPath: /models
            - name: local-image-tmp
              mountPath: /tmp
          readinessProbe:
            httpGet:
              path: /health
              port: local-image
            initialDelaySeconds: 10
            periodSeconds: 10
            timeoutSeconds: 5
          livenessProbe:
            httpGet:
              path: /health
              port: local-image
            initialDelaySeconds: 30
            periodSeconds: 30
            timeoutSeconds: 5
          securityContext:
            allowPrivilegeEscalation: false
            capabilities:
              drop: [ALL]
            readOnlyRootFilesystem: true
            seccompProfile:
              type: RuntimeDefault
          resources:
            requests:
              cpu: "4"
              memory: 12Gi
              nvidia.com/gpu.shared: 1
            limits:
              cpu: "12"
              memory: 28Gi
              nvidia.com/gpu.shared: 1