Update llama runtime and isolate background reviews

compose.yaml

@@ -32,6 +32,76 @@ x-llama-common: &llama-common
     start_period: 30s
 
 services:
+  # Kleines, dauerhaft geladenes Hilfsmodell fuer Hermes-Hintergrundaufgaben
+  # (Memory-/Skill-Review). Es laeuft unabhaengig vom wechselnden Hauptmodell
+  # und kann dessen Prompt-Cache daher nicht mehr
+  # verdraengen. Athena reserviert die 3060 bereits fuer Qwen und XTTS; der
+  # robuste Standard ist deshalb CPU/RAM. GPU-Layer koennen spaeter per Env
+  # zugeschaltet werden, sobald dort garantiert Speicher frei ist.
+  llama-review:
+    image: ${LLAMA_IMAGE:-mike-ai/llama.cpp:local}
+    container_name: mike-ai-llama-review
+    restart: unless-stopped
+    gpus: all
+    ipc: host
+    read_only: true
+    tmpfs:
+      - /tmp:size=512m,mode=1777
+    volumes:
+      - "${MODEL_DIR:-/srv/mike-ai/models}:/models:ro"
+    environment:
+      NVIDIA_VISIBLE_DEVICES: ${REVIEW_GPU_DEVICE:-1}
+      NVIDIA_DRIVER_CAPABILITIES: compute,utility
+    command:
+      - --model
+      - "/models/${REVIEW_MODEL_FILE:-qwen3-4b-review/Qwen3-4B-Instruct-2507-Q4_K_M.gguf}"
+      - --alias
+      - qwen-review
+      - --ctx-size
+      - "${REVIEW_CONTEXT:-32768}"
+      - --flash-attn
+      - "on"
+      - --cache-type-k
+      - q4_0
+      - --cache-type-v
+      - q4_0
+      - --cache-prompt
+      - --threads
+      - "${LLAMA_THREADS:-6}"
+      - --threads-batch
+      - "${LLAMA_THREADS_BATCH:-6}"
+      - --batch-size
+      - "${REVIEW_BATCH_SIZE:-512}"
+      - --ubatch-size
+      - "${REVIEW_UBATCH_SIZE:-64}"
+      - --parallel
+      - "1"
+      - --kv-unified
+      - --jinja
+      - --host
+      - 0.0.0.0
+      - --port
+      - "8080"
+      - --metrics
+      - --fit
+      - "off"
+      - --n-gpu-layers
+      - "${REVIEW_GPU_LAYERS:-0}"
+      - --no-kv-offload
+      - --no-mmap
+      - --no-ui
+      - --temperature
+      - "0.2"
+    networks: [inference]
+    security_opt: ["no-new-privileges:true"]
+    cap_drop: [ALL]
+    healthcheck:
+      test: [CMD, curl, -fsS, "http://127.0.0.1:8080/health"]
+      interval: 10s
+      timeout: 5s
+      retries: 60
+      start_period: 30s
+
   wireguard-gateway:
     build: ./platform/docker/wireguard-gateway
     image: mike-ai/wireguard-gateway:local
@@ -153,17 +223,12 @@ services:
     environment:
       NVIDIA_VISIBLE_DEVICES: ${MEDIUM_GPU_DEVICES:-0,1}
       NVIDIA_DRIVER_CAPABILITIES: compute,utility
-      # Keep the language model split unchanged while placing the complete
-      # multimodal projector on the secondary RTX 3060.
-      MTMD_BACKEND_DEVICE: CUDA1
+      # Temporary llama.cpp cache workaround: keep Medium text-only. Current
+      # multimodal builds can discard the complete KV state on a later turn.
+      # The model split, context and MTP settings remain unchanged.
     command:
       - --model
       - "/models/${MEDIUM_MODEL_FILE:?MEDIUM_MODEL_FILE is required}"
-      - --mmproj
-      - "/models/${VISION_PROJECTOR_FILE:?VISION_PROJECTOR_FILE is required}"
-      - --mmproj-offload
-      - --mmproj-device
-      - CUDA1
       - --alias
       - qwen-medium
       - --ctx-size
@@ -568,6 +633,9 @@ services:
       ROUTER_STATE_FILE: /var/lib/mike-ai-profile-router/state.json
       ROUTER_MAX_CONCURRENT_REQUESTS: "16"
       UPSTREAM_URL: http://llama-upstream:8080
+      REVIEW_UPSTREAM_URL: http://llama-review:8080
+      REVIEW_MODEL_NAME: qwen-review
+      REVIEW_CONTEXT_LENGTH: "${REVIEW_CONTEXT:-32768}"
       PROFILE_CONTROL_URL: http://profile-controller:8090
       PROFILE_CONTROL_TOKEN: "${CONTROLLER_TOKEN:?CONTROLLER_TOKEN is required}"
       SWITCH_TIMEOUT: "600"