TTS-Slots (Piper CPU, Qwen3-TTS auf Ada), M4000 außer Betrieb, Engine-Grenzen im Placement (ADR-0006)

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

deploy/compose/gpu-vllm/.env.example

@@ -7,3 +7,6 @@ MAX_LORA_RANK=64
 MODELS_DIR=/srv/models
 ADAPTERS_DIR=/srv/adapters
 HF_TOKEN=
+# nur gpu-flex mit Profil "tts": GPU_MEM_UTIL für vLLM dann auf ~0.55 senken, damit Qwen3-TTS Platz hat
+QWEN_TTS_IMAGE=
+QWEN_TTS_MODEL=Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice

deploy/compose/gpu-vllm/docker-compose.yml

@@ -36,6 +36,24 @@ services:
       retries: 3
       start_period: 300s   # Laden großer Modelle
 
+  # Qwen3-TTS (ADR-0006) – nur auf gpu-flex, Profil "tts". Server muss OpenAI-kompatibles
+  # /v1/audio/speech anbieten (z. B. ein Wrapper um faster-qwen3-tts); Image-Wahl beim Deployment.
+  qwen-tts:
+    image: ${QWEN_TTS_IMAGE:-knaster/qwen-tts:unset}
+    container_name: qwen-tts
+    profiles: ["tts"]
+    restart: unless-stopped
+    labels: {autoheal: "true"}
+    environment:
+      TTS_MODEL: ${QWEN_TTS_MODEL:-Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice}
+    volumes:
+      - ${MODELS_DIR:-/srv/models}:/root/.cache/huggingface
+    ports: ["8091:8091"]
+    deploy:
+      resources:
+        reservations:
+          devices: [{driver: nvidia, count: 1, capabilities: [gpu]}]
+
   gpu-exporter:
     image: utkuozdemir/nvidia_gpu_exporter:1.2.1
     restart: unless-stopped

deploy/compose/infra/docker-compose.yml

@@ -37,6 +37,20 @@ services:
       timeout: 5s
       retries: 3
 
+  tts-piper:
+    <<: *restart
+    build:
+      context: ../../..
+      dockerfile: deploy/docker/tts-piper.Dockerfile
+    image: knaster/tts-piper:${KNASTER_VERSION:-dev}
+    container_name: tts-piper
+    environment:
+      KNASTER_TTS_PARALLEL: ${TTS_PARALLEL:-2}   # i7 mit 4 Kernen: max. 2 parallele Synthesen
+    ports: ["8090:8090"]
+    deploy:
+      resources:
+        limits: {cpus: "2", memory: 1G}
+
   postgres:
     <<: *restart
     image: pgvector/pgvector:pg16