gpu(titan-24): trust checkout lease over desktop spikes

services/hermes/local-image-deployment.yaml

@@ -67,10 +67,11 @@ spec:
               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
+            # The lease is authoritative for checkout state. Keep a high SM
-            # 3% SM. Active desktop/game sessions cross one of these bounds.
+            # 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: "8"
+              value: "80"
             - name: HERMES_LOCAL_IMAGE_GPU_MAX_EXTERNAL_MEMORY_BYTES
               value: "3221225472"
             - name: HERMES_LOCAL_IMAGE_GPU_ACTIVITY_SAMPLES

testing/tests/test_hermes_chat_quality.py

@@ -621,7 +621,7 @@ def test_local_flux_runtime_and_gpu_handoff_are_flux_managed():
         "e7b7dc27f91deacad38e78976d1f2b499d76a294"
     )
     assert model_env["HERMES_LOCAL_IMAGE_GPU_ACTIVITY_NODE"] == "titan-24"
-    assert model_env["HERMES_LOCAL_IMAGE_GPU_ACTIVE_SM_PERCENT"] == "8"
+    assert model_env["HERMES_LOCAL_IMAGE_GPU_ACTIVE_SM_PERCENT"] == "80"
     assert model_env["HERMES_LOCAL_IMAGE_GPU_MAX_EXTERNAL_MEMORY_BYTES"] == (
         "3221225472"
     )
@@ -692,15 +692,15 @@ def test_titan20_serializes_classifier_and_local_chat_model_residency():
     )
 
 
-def test_local_image_gpu_guard_distinguishes_idle_and_active_desktop(monkeypatch):
+def test_local_image_gpu_guard_distinguishes_background_and_saturated_gpu(monkeypatch):
-    """Idle display shells may coexist, but active desktop/game work blocks FLUX."""
+    """Lease-idle desktop spikes may coexist, but saturation still blocks FLUX."""
     source = ROOT / "dockerfiles" / "hermes-local-image-server.py"
     spec = importlib.util.spec_from_file_location("hermes_local_image_server", source)
     assert spec and spec.loader
     module = importlib.util.module_from_spec(spec)
     spec.loader.exec_module(module)
     monkeypatch.setattr(module, "GPU_ACTIVITY_NODE", "titan-24")
-    monkeypatch.setattr(module, "GPU_ACTIVE_SM_PERCENT", 8.0)
+    monkeypatch.setattr(module, "GPU_ACTIVE_SM_PERCENT", 80.0)
     monkeypatch.setattr(module, "GPU_MAX_EXTERNAL_MEMORY_BYTES", 3 << 30)
 
     idle = module._parse_gpu_activity(
@@ -717,11 +717,16 @@ def test_local_image_gpu_guard_distinguishes_idle_and_active_desktop(monkeypatch
     assert idle["external_gpu_memory_bytes"] == 1900000000
     assert idle["external_gpu_sm_percent"] == 3
 
+    background_spike = module._parse_gpu_activity(
+        'nvidia_process_gpu_sm_util_percent{node="titan-24",namespace="host",process="sway"} 41\n'
+    )
+    assert background_spike["interactive_active"] is False
+
     active = module._parse_gpu_activity(
-        'nvidia_process_gpu_sm_util_percent{node="titan-24",namespace="host",process="steam"} 41\n'
+        'nvidia_process_gpu_sm_util_percent{node="titan-24",namespace="host",process="steam"} 91\n'
     )
     assert active["interactive_active"] is True
-    assert "41%" in active["gpu_guard_reason"]
+    assert "91%" in active["gpu_guard_reason"]
 
 
 def test_local_flux_renderer_uses_a_disposable_cuda_worker(monkeypatch):