#!/bin/bash
# Qwen3-TTS 1.7B auf der DGX Spark (admin@10.146.97.103) via vllm-omni.
# OpenAI-kompatibel: POST http://10.146.97.103:<port>/v1/audio/speech
#
#   ./start-qwen3-tts.sh             -> CustomVoice (9 feste Stimmen)  :8011
#   ./start-qwen3-tts.sh base        -> Base (Voice-Cloning, ref_audio) :8012
#
# Liegt auf der DGX unter ~/qwen3-tts/ zusammen mit qwen3_tts_spark.yaml
# (eigene Deploy-Config: V1-Runner + 0.08/0.06 Speicher, siehe Kopf der YAML).
set -eu
CFG_DIR=/home/admin/qwen3-tts
IMAGE=vllm-node-030-omni   # vllm-omni 0.30.0 + librosa/soundfile, bringt qwen3_tts mit

case "${1:-customvoice}" in
  customvoice) NAME=qwen3-tts-customvoice; PORT=8011; MODEL=Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice ;;
  base)        NAME=qwen3-tts-base;        PORT=8012; MODEL=Qwen/Qwen3-TTS-12Hz-1.7B-Base ;;
  *) echo "Aufruf: $0 [customvoice|base]" >&2; exit 1 ;;
esac

if docker ps --format '{{.Names}}' | grep -qx "$NAME"; then
  echo "$NAME laeuft bereits"; exit 0
fi
docker rm "$NAME" >/dev/null 2>&1 || true

docker run -d --name "$NAME" --network host --ipc host --gpus all \
  -v /home/admin/.cache/huggingface:/root/.cache/huggingface \
  -v "$CFG_DIR":/cfg:ro \
  -e VLLM_DISABLE_COMPILE_CACHE=1 \
  -e HF_HUB_OFFLINE=1 \
  "$IMAGE" \
  vllm-omni serve "$MODEL" --omni --port "$PORT" \
    --deploy-config /cfg/qwen3_tts_spark.yaml

echo "gestartet: docker logs -f $NAME"