gpu(titan-24): pin image guard to local exporter

services/hermes/local-image-deployment.yaml

@@ -64,7 +64,7 @@ spec:
             - name: LEASE_IMAGE_OWNER
               value: hermes-image
             - name: HERMES_LOCAL_IMAGE_GPU_ACTIVITY_URL
-              value: http://nvidia-process-exporter.monitoring.svc.cluster.local:9401/metrics
+              value: http://nvidia-process-exporter-local.monitoring.svc.cluster.local:9401/metrics
             - name: HERMES_LOCAL_IMAGE_GPU_ACTIVITY_NODE
               value: titan-24
             # The idle desktop stack and Wolf daemon use about 1.9 GiB and

services/monitoring/nvidia-process-exporter.yaml

@@ -141,3 +141,21 @@ spec:
     - name: metrics
       port: 9401
       targetPort: metrics
+---
+apiVersion: v1
+kind: Service
+metadata:
+  name: nvidia-process-exporter-local
+  namespace: monitoring
+  labels:
+    app: nvidia-process-exporter
+spec:
+  # GPU arbitration must inspect the exporter on its own node. The regular
+  # service intentionally balances all exporters for cluster monitoring.
+  internalTrafficPolicy: Local
+  selector:
+    app: nvidia-process-exporter
+  ports:
+    - name: metrics
+      port: 9401
+      targetPort: metrics

testing/tests/test_hermes_chat_quality.py

@@ -613,7 +613,7 @@ def test_local_flux_runtime_and_gpu_handoff_are_flux_managed():
     assert model_env["HERMES_LOCAL_IMAGE_GPU_MAX_EXTERNAL_MEMORY_BYTES"] == (
         "3221225472"
     )
-    assert "nvidia-process-exporter.monitoring.svc.cluster.local" in model_env[
+    assert "nvidia-process-exporter-local.monitoring.svc.cluster.local" in model_env[
         "HERMES_LOCAL_IMAGE_GPU_ACTIVITY_URL"
     ]
     models_volume = next(item for item in pod["volumes"] if item["name"] == "models")