feat(observability): add DCGM nvidia-exporter for GPU metrics

compose.yml

@@ -599,6 +599,25 @@ services:
       start_period: 30s
     restart: unless-stopped
 
+  nvidia-exporter:
+    image: nvcr.io/nvidia/k8s/dcgm-exporter:3.3.9-3.6.1-ubuntu22.04
+    container_name: nvidia-exporter
+    networks:
+      - observability
+    runtime: nvidia
+    environment:
+      - DCGM_EXPORTER_NO_HOSTNAME=1
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: 1
+              capabilities: [gpu]
+    cap_add:
+      - SYS_ADMIN
+    restart: unless-stopped
+
   promtail:
     image: grafana/promtail:latest
     container_name: promtail

tests/test_observability.py

@@ -59,3 +59,18 @@ class TestCadvisor:
         # rootfs, sys, /dev/disk needed for resource introspection
         assert any(s == "/" for s in sources)
         assert any(s == "/sys" for s in sources)
+
+
+class TestNvidiaExporter:
+    def test_nvidia_exporter_present(self):
+        svcs = _services()
+        assert "nvidia-exporter" in svcs
+
+    def test_nvidia_runtime(self):
+        svc = _services()["nvidia-exporter"]
+        assert svc.get("runtime") == "nvidia"
+
+    def test_nvidia_gpu_reservation(self):
+        svc = _services()["nvidia-exporter"]
+        devices = svc["deploy"]["resources"]["reservations"]["devices"]
+        assert any(d.get("driver") == "nvidia" for d in devices)