hermes: stabilize local GPU workloads

dockerfiles/hermes-local-image-server.py

@@ -19,7 +19,9 @@ from urllib.request import Request, urlopen
 
 
 HOST = os.environ.get("HERMES_LOCAL_IMAGE_HOST", "0.0.0.0")
-PORT = int(os.environ.get("HERMES_LOCAL_IMAGE_PORT", "9004"))
+# Kubernetes reserves <SERVICE>_PORT for service-link discovery. Keep the
+# listener setting out of that namespace so it is always numeric.
+PORT = int(os.environ.get("HERMES_LOCAL_IMAGE_LISTEN_PORT", "9004"))
 MODEL_ID = os.environ.get(
     "HERMES_LOCAL_IMAGE_MODEL", "black-forest-labs/FLUX.2-klein-4B"
 )

services/ai-llm/deployment.yaml

@@ -129,7 +129,7 @@ spec:
               pid="$!"
               trap 'kill -TERM "$pid"; wait "$pid"' TERM INT
               sleep 6
-              if ! timeout 180s ollama run "${OLLAMA_FAST_MODEL}" --keepalive=-1 "reply with just pong" >/tmp/ollama-fast-warm.log 2>&1; then
+              if ! timeout 180s ollama run "${OLLAMA_FAST_MODEL}" "reply with just pong" >/tmp/ollama-fast-warm.log 2>&1; then
                 cat /tmp/ollama-fast-warm.log >&2
                 exit 1
               fi

services/hermes/local-image-deployment.yaml

@@ -43,7 +43,7 @@ spec:
             sizeLimit: 2Gi
       containers:
         - name: local-image
-          image: registry.bstein.dev/bstein/hermes-local-image@sha256:2bf0e5b19ef80cf0cfe7b3d9a4b672df101a3215634a6cede85ff850ada7a28a
+          image: registry.bstein.dev/bstein/hermes-local-image@sha256:316d123cc48509a53fb32c26bfb384a3ed4d934cb334d73faa0840a62e047e65
           imagePullPolicy: IfNotPresent
           ports:
             - name: local-image
@@ -53,6 +53,8 @@ spec:
               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

testing/tests/test_hermes_chat_quality.py

@@ -604,6 +604,7 @@ def test_local_flux_runtime_and_gpu_handoff_are_flux_managed():
     assert local["ports"] == [{"name": "local-image", "containerPort": 9004}]
     assert any(mount["mountPath"] == "/models" for mount in local["volumeMounts"])
     model_env = {item["name"]: item["value"] for item in local["env"]}
+    assert model_env["HERMES_LOCAL_IMAGE_LISTEN_PORT"] == "9004"
     assert model_env["HERMES_LOCAL_IMAGE_REVISION"] == (
         "e7b7dc27f91deacad38e78976d1f2b499d76a294"
     )
@@ -663,6 +664,8 @@ def test_titan20_serializes_classifier_and_local_chat_model_residency():
     assert env["OLLAMA_NUM_PARALLEL"] == "1"
     assert env["OLLAMA_KEEP_ALIVE"] == "-1"
     assert env["OLLAMA_CONTEXT_LENGTH"] == "8192"
+    warm_command = " ".join(container["command"])
+    assert "--keepalive=-1" not in warm_command
     models = next(item for item in pod["volumes"] if item["name"] == "models")
     assert models["persistentVolumeClaim"]["claimName"] == (
         "ollama-models-titan20"