fix(ai): use pod-local ollama model cache

clusters/atlas/flux-system/applications/ai-llm/kustomization.yaml

@@ -8,7 +8,8 @@ metadata:
     kustomize.toolkit.fluxcd.io/ssa: IfNotPresent
 spec:
   interval: 10m
-  suspend: true
+  suspend: false
+  timeout: 30m
   path: ./services/ai-llm
   targetNamespace: ai
   prune: true

services/ai-llm/deployment.yaml

@@ -37,8 +37,8 @@ spec:
       runtimeClassName: nvidia
       volumes:
         - name: models
-          persistentVolumeClaim:
-            claimName: ollama-models-asteria
+          # Accelerator nodes are not Longhorn attach targets; keep the model cache pod-local.
+          emptyDir: {}
       initContainers:
         - name: warm-model
           image: ollama/ollama@sha256:2c9595c555fd70a28363489ac03bd5bf9e7c5bdf2890373c3a830ffd7252ce6d
@@ -95,6 +95,13 @@ spec:
           volumeMounts:
             - name: models
               mountPath: /root/.ollama
+          readinessProbe:
+            httpGet:
+              path: /api/tags
+              port: 11434
+            initialDelaySeconds: 15
+            periodSeconds: 10
+            timeoutSeconds: 5
           resources:
             requests:
               cpu: "4"