Add a busy timeout for the LocalAI watchdog

LOCALAI_WATCHDOG_BUSY had no matching timeout setting, so the role
could only ever use LocalAI's built-in 5 minutes. On the throttled
Tesla T4 a legitimate request can take more than 6 minutes, so default
to 10 minutes, which also matches the ingress proxy-read-timeout.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

defaults/main.yml

@@ -8,3 +8,17 @@ localai_api_keys:
 localai_watchdog_idle: "false"
 localai_watchdog_idle_timeout: "15m"
 localai_watchdog_busy: "true"
+localai_watchdog_busy_timeout: "10m"
+localai_image_tag: latest-gpu-nvidia-cuda-13
+localai_debug: "false"
+localai_p2p: "false"
+localai_federated: "false"
+# Persistent state lives on the models PVC (mounted at /models).
+localai_data_path: /models/.localai/data
+localai_config_dir: /models/.localai/configuration
+localai_backends_path: /models/.backends
+localai_agent_pool_default_model: ornith-1.5-9b-q8
+localai_agent_pool_embedding_model: granite-embedding-107m-multilingual
+localai_host_network: true
+# Name of a Secret holding LOCALAI_P2P_TOKEN; empty skips the patch.
+localai_p2p_token_secret: localai-p2p

tasks/main.yml

@@ -20,7 +20,7 @@
     values:
       deployment:
         image:
-          tag: latest-gpu-nvidia-cuda-12
+          tag: "{{ localai_image_tag }}"
         pullPolicy: Always
         runtimeClassName: nvidia
         secretEnv:
@@ -34,12 +34,27 @@
             value: "{{ localai_watchdog_idle_timeout }}"
           - name: LOCALAI_WATCHDOG_BUSY
             value: "{{ localai_watchdog_busy }}"
+          - name: LOCALAI_WATCHDOG_BUSY_TIMEOUT
+            value: "{{ localai_watchdog_busy_timeout }}"
           - name: DEBUG
             value: "{{ localai_debug }}"
           - name: LOCALAI_P2P
             value: "{{ localai_p2p }}"
           - name: LOCALAI_FEDERATED
             value: "{{ localai_federated }}"
+          # Keep agents, knowledge-base collections, users and runtime settings
+          # on the models PVC; the chart's default /data and /configuration are
+          # ephemeral, so a pod restart used to wipe them.
+          - name: LOCALAI_DATA_PATH
+            value: "{{ localai_data_path }}"
+          - name: LOCALAI_CONFIG_DIR
+            value: "{{ localai_config_dir }}"
+          - name: LOCALAI_BACKENDS_PATH
+            value: "{{ localai_backends_path }}"
+          - name: LOCALAI_AGENT_POOL_DEFAULT_MODEL
+            value: "{{ localai_agent_pool_default_model }}"
+          - name: LOCALAI_AGENT_POOL_EMBEDDING_MODEL
+            value: "{{ localai_agent_pool_embedding_model }}"
       resources:
         limits:
           nvidia.com/gpu: 1
@@ -70,3 +85,39 @@
             hosts:
               - localai.eom.dev
 
+# The chart has no values for these, so patch them onto the Deployment after
+# Helm. hostNetwork lets p2p workers on the LAN reach the server directly.
+- name: Enable host networking
+  kubernetes.core.k8s:
+    state: patched
+    kind: Deployment
+    name: localai-local-ai
+    namespace: localai
+    definition:
+      spec:
+        template:
+          spec:
+            hostNetwork: true
+            dnsPolicy: ClusterFirstWithHostNet
+  when: localai_host_network | bool
+
+# The p2p token lives in a Secret created outside this role (localai/localai-p2p).
+- name: Read the p2p token from its Secret
+  kubernetes.core.k8s:
+    state: patched
+    kind: Deployment
+    name: localai-local-ai
+    namespace: localai
+    definition:
+      spec:
+        template:
+          spec:
+            containers:
+              - name: local-ai
+                env:
+                  - name: LOCALAI_P2P_TOKEN
+                    valueFrom:
+                      secretKeyRef:
+                        name: "{{ localai_p2p_token_secret }}"
+                        key: LOCALAI_P2P_TOKEN
+  when: localai_p2p_token_secret | length > 0