Remove obsolete review model service

.env.example

@@ -1,10 +1,6 @@
 # Generated as /etc/mike-ai/stack.env by install.sh. Never commit real values.
 AI_BIND_ADDRESS=10.77.0.2
 MODEL_DIR=/data/models
-REVIEW_MODEL_FILE=qwen3-4b-review/Qwen3-4B-Instruct-2507-Q4_K_M.gguf
-REVIEW_CONTEXT=32768
-REVIEW_GPU_DEVICE=1
-REVIEW_GPU_LAYERS=0
 ROUTER_API_KEY=GENERATED_BY_INSTALLER
 CONTROLLER_TOKEN=GENERATED_BY_INSTALLER
 FLUX_MODEL_DIR=/data/models/FLUX.2-klein-4B

compose.yaml

@@ -32,76 +32,6 @@ 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
@@ -634,9 +564,6 @@ 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"

config/install.env.example

@@ -4,12 +4,6 @@
 AI_HOSTNAME=ki-host
 ADMIN_USER=mike
 MODEL_DIR=/data/models
-REVIEW_MODEL_FILE=qwen3-4b-review/Qwen3-4B-Instruct-2507-Q4_K_M.gguf
-REVIEW_CONTEXT=32768
-# Eine einzelne GPU-ID oder stabile NVIDIA-GPU-UUID; auf Athena die RTX 3060.
-REVIEW_GPU_DEVICE=1
-REVIEW_GPU_LAYERS=0
-
 # Installing a new NVIDIA driver can require one reboot. In that case this
 # installer exits with code 20 (NVIDIA) or 21 (stable NIC rename); rerun the
 # same command after reboot.

docs/CURRENT_RUNTIME_NOTES.md

@@ -4,7 +4,7 @@ Stand: 31. August 2026
 
 ## Produktive llama.cpp-Runtime
 
-Alle Textprofile und das Review-Modell verwenden llama.cpp Build 10718,
+Alle Textprofile verwenden llama.cpp Build 10718,
 Commit `41ef91f7c8046087cdfbb276b79bff311ecf1c6d`. Der vorherige Build 10587
 wurde nach einem kontrollierten A/B-Test ersetzt. Der Wechsel verbessert den
 Durchsatz bei großen Werkzeuglisten, ohne im 70K-/120K-Kontext- und Cachetest