RTX 4070 Ti 自宅サーバーに Qwen3-TTS を入れた実録 — API変更で3回詰まった話

mkdir -p ~/qwen3-tts/output
chmod 700 ~/qwen3-tts
cd ~/qwen3-tts
python3 -m venv .venv
source .venv/bin/activate
pip install --upgrade pip

pip install torch torchaudio \
    --index-url https://download.pytorch.org/whl/cu124

pip install qwen-tts
sudo apt install sox

# api_server.py
from contextlib import asynccontextmanager
import asyncio, io, os
import soundfile as sf
import torch
from fastapi import FastAPI, HTTPException
from fastapi.responses import Response
from pydantic import BaseModel, Field
from typing import Optional
from dotenv import load_dotenv

load_dotenv()

MODEL_NAME = os.getenv("MODEL_NAME", "Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice")
_model = None
_gpu_lock = asyncio.Lock()

@asynccontextmanager
async def lifespan(app: FastAPI):
    global _model
    from qwen_tts import Qwen3TTSModel  # ★ QwenTTS ではない
    _model = Qwen3TTSModel.from_pretrained(
        MODEL_NAME,
        device_map="cuda:0",
        dtype=torch.bfloat16,       # ★ torch_dtype= ではなく dtype=
        attn_implementation="sdpa", # FlashAttention2 未インストール時はこれ
    )
    yield

app = FastAPI(lifespan=lifespan)

class SynthesizeRequest(BaseModel):
    text: str = Field(..., min_length=1, max_length=2000)
    speaker: str = Field(default="Ono_Anna")
    language: str = Field(default="Japanese")
    instruct: Optional[str] = None

@app.get("/health")
async def health():
    return {"status": "ok", "model_loaded": _model is not None}

@app.post("/v1/tts/synthesize")
async def synthesize(body: SynthesizeRequest):
    if _model is None:
        raise HTTPException(503, "モデルロード中")
    async with _gpu_lock:
        loop = asyncio.get_event_loop()
        wavs, sr = await loop.run_in_executor(
            None,
            lambda: _model.generate_custom_voice( # ★ synthesize() ではない
                text=body.text,
                speaker=body.speaker,              # ★ speaker_name= ではなく speaker=
                language=body.language,
                instruct=body.instruct,
            )
        )
    buf = io.BytesIO()
    sf.write(buf, wavs[0], sr, format="WAV")  # ★ 戻り値は (List[ndarray], int) → wavs[0]
    return Response(content=buf.getvalue(), media_type="audio/wav")

# /etc/systemd/system/qwen3tts.service
[Unit]
Description=Qwen3-TTS Speech Synthesis Server
After=network-online.target
Conflicts=comfyui.service acestep.service

[Service]
Type=simple
User=YOUR_USER
WorkingDirectory=/home/YOUR_USER/qwen3-tts
ExecStart=/home/YOUR_USER/qwen3-tts/.venv/bin/uvicorn \
    api_server:app --host 0.0.0.0 --port 9001 --workers 1
Restart=on-failure
EnvironmentFile=/home/YOUR_USER/qwen3-tts/.env
NoNewPrivileges=true
ProtectSystem=strict
ReadWritePaths=/home/YOUR_USER/qwen3-tts /home/YOUR_USER/.cache/huggingface
MemoryMax=16G
TimeoutStartSec=300

[Install]
WantedBy=multi-user.target

sudo systemctl daemon-reload
sudo systemctl start qwen3tts
journalctl -u qwen3tts -f

sudo ufw allow from 192.168.x.x to any port 9001 proto tcp

curl -X POST http://localhost:9001/v1/tts/synthesize \
  -H "Content-Type: application/json" \
  -d '{"text":"こんにちは、テスト音声です。"}' \
  -o test.wav && file test.wav

ImportError: cannot import name 'QwenTTS' from 'qwen_tts'

import qwen_tts; print(dir(qwen_tts))
# → ['Qwen3TTSModel', 'Qwen3TTSTokenizer', 'VoiceClonePromptItem', ...]

# ❌
Qwen3TTSModel.from_pretrained(MODEL, torch_dtype=torch.bfloat16)

# ✅
Qwen3TTSModel.from_pretrained(MODEL, dtype=torch.bfloat16)

try:
    model = Qwen3TTSModel.from_pretrained(MODEL, attn_implementation="flash_attention_2", ...)
except Exception:
    model = Qwen3TTSModel.from_pretrained(MODEL, attn_implementation="sdpa", ...)

if ! sudo cat /etc/systemd/system/foo.service | grep -q "bar"; then
    if sudo cat /etc/systemd/system/foo.service | grep -q "^Conflicts="; then
        sudo sed -i '/^Conflicts=/ s/$/ bar.service/' /etc/systemd/system/foo.service
    else
        sudo sed -i '/^\[Unit\]/a Conflicts=bar.service' /etc/systemd/system/foo.service
    fi
fi

[Unit]
Conflicts=comfyui.service acestep.service