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

deploy/docker/tts-piper.Dockerfile

@@ -0,0 +1,30 @@
+# Piper-TTS hinter dem OpenAI-kompatiblen knaster-TTS-Dienst. Build-Kontext: Repo-Wurzel.
+# Hinweis: piper-tts ab 1.3 ist GPL-3.0 – läuft hier als eigenständiger Dienst.
+FROM python:3.12-slim
+
+ARG PIPER_VERSION=1.3.0
+ARG VOICE_BASE=https://huggingface.co/rhasspy/piper-voices/resolve/main/de/de_DE
+ENV PYTHONDONTWRITEBYTECODE=1 PYTHONUNBUFFERED=1 PIP_NO_CACHE_DIR=1
+WORKDIR /app
+
+RUN apt-get update -qq && apt-get install -y -qq --no-install-recommends curl && rm -rf /var/lib/apt/lists/*
+RUN pip install "piper-tts==${PIPER_VERSION}"
+
+COPY core/pyproject.toml core/pyproject.toml
+COPY core/src core/src
+RUN pip install ./core
+
+RUN mkdir -p /voices && cd /voices \
+ && curl -fsSL -o de_DE-thorsten-medium.onnx      "${VOICE_BASE}/thorsten/medium/de_DE-thorsten-medium.onnx" \
+ && curl -fsSL -o de_DE-thorsten-medium.onnx.json "${VOICE_BASE}/thorsten/medium/de_DE-thorsten-medium.onnx.json"
+
+ENV KNASTER_TTS_VOICES=thorsten=/voices/de_DE-thorsten-medium.onnx \
+    KNASTER_TTS_CMD="piper -m {model} -f {out}" \
+    KNASTER_TTS_PARALLEL=2
+
+RUN useradd --system --uid 10002 tts
+USER tts
+EXPOSE 8090
+HEALTHCHECK --interval=30s --timeout=5s --retries=3 \
+  CMD python -c "import urllib.request; urllib.request.urlopen('http://127.0.0.1:8090/health', timeout=4)"
+CMD ["knaster", "tts-serve", "--host", "0.0.0.0", "--port", "8090"]

deploy/litellm/config.yaml

@@ -58,10 +58,25 @@ model_list:
       api_key: none
     model_info: {mode: rerank}
 
+  # --- Sprachausgabe (OpenAI-kompatibles /v1/audio/speech) ---
+  - model_name: tts-de                              # Piper im Infra-LXC (knaster tts-serve)
+    litellm_params:
+      model: openai/tts-de
+      api_base: http://tts-piper:8090/v1
+      api_key: none
+    model_info: {mode: audio_speech}
+  - model_name: tts-de-hq                           # Qwen3-TTS auf gpu-flex (Server-Wahl beim Deployment)
+    litellm_params:
+      model: openai/tts-de-hq
+      api_base: http://gpu-flex:8091/v1
+      api_key: none
+    model_info: {mode: audio_speech}
+
 router_settings:
   # Infrastruktur-Fallbacks innerhalb einer Stufe; Qualitäts-Eskalation macht der Router-Service
   fallbacks:
     - tier1-general: [tier1-flex]
+    - tts-de-hq: [tts-de]
   allowed_fails: 2
   cooldown_time: 60
   num_retries: 1