Hidosi/umb
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: preload Whisper model at startup with disk space check

Browse Source 
- Add WHISPER_MODEL_SIZE and WHISPER_MIN_FREE_SPACE_BYTES config options
- voice.py now preloads/downloads the Whisper model at bot startup
- Detailed logging for model presence, disk space, download progress and readiness
- If model is already present locally, preload is skipped
- If disk space is insufficient, bot fails fast with a clear error
- main.py calls preload_model() after proxy setup and before polling
- Document voice recognition model download behavior in README
This commit is contained in:
Галингер Р.С.
2026-07-08 09:44:48 +07:00
parent 0432155c8f
commit d6c029f708
4 changed files with 137 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files
README.md
+43
View File
@@ -95,6 +95,10 @@ ROUTERAI_BASE_URL=https://routerai.ru/api/v1
# Внимание: проверка расходует токены API. По умолчанию отключена.
AI_HEALTH_CHECK_ENABLED=false
AI_HEALTH_CHECK_INTERVAL=600
# Настройки модели Whisper для распознавания голоса.
WHISPER_MODEL_SIZE=base
WHISPER_MIN_FREE_SPACE_BYTES=5368709120
```
Для включения прокси установите `PROXY_ENABLED=true` и укажите корректный `PROXY_URL`.
@@ -114,6 +118,45 @@ docker compose up -d --build
Модели Whisper сохраняются в `./models`, база данных и логи — в `./bot/data`.
## 🎙️ Распознавание голосовых сообщений
Бот автоматически распознаёт все голосовые сообщения с помощью локальной модели [faster-whisper](https://github.com/SYSTRAN/faster-whisper).
### Поведение загрузки модели
- При запуске бота проверяется наличие модели в `./models/faster-whisper/<WHISPER_MODEL_SIZE>/`.
- Если модель уже скачана — бот сразу готов к работе.
- Если модели нет — бот проверяет свободное место на диске и **скачивает модель автоматически** перед началом polling.
- Для скачивания используется прокси из `.env`, если `PROXY_ENABLED=true`.
### Требования к месту
| Модель | Размер модели | Рекомендуемое свободное место |
|--------|---------------|-------------------------------|
| `base` | ~150 MB архив, ~500 MB на диске | 5 GB (`WHISPER_MIN_FREE_SPACE_BYTES=5368709120`) |
| `small` | ~500 MB архив, ~1.5 GB на диске | 5 GB+ |
| `medium` | ~1.5 GB архив, ~5 GB на диске | 10 GB+ |
> 💡 По умолчанию используется модель `base`. Для смены модели измените `WHISPER_MODEL_SIZE` в `.env`.
### Логирование
В логах будет видно:
```
Free disk space: 45.23 GB, required: 5.00 GB
Whisper model 'base' found locally at /app/models/faster-whisper/base
Whisper model 'base' is ready
```
или при скачивании:
```
Whisper model 'base' not found locally. Starting download to /app/models/faster-whisper...
Whisper model 'base' downloaded and loaded in 125.4s
Whisper model 'base' is ready
```
## 🧪 Тесты
```bash
bot/utils/voice.py
+80 -7
View File
@@ -2,12 +2,14 @@ import asyncio
import logging
import os
import re
import shutil
import time
from pathlib import Path
from faster_whisper import WhisperModel
from aiogram import Bot
from config import PROXY_ENABLED, PROXY_URL
from config import PROXY_ENABLED, PROXY_URL, WHISPER_MODEL_SIZE, WHISPER_MIN_FREE_SPACE_BYTES
logger = logging.getLogger(__name__)
@@ -25,26 +27,59 @@ def _setup_proxy() -> None:
logger.info("Proxy set for model download: %s...", PROXY_URL[:30])
def _get_model_path() -> Path:
return MODELS_DIR / WHISPER_MODEL_SIZE
def _is_model_downloaded() -> bool:
"""Check whether the Whisper model files are already present locally."""
model_path = _get_model_path()
model_bin = model_path / "model.bin"
return model_path.is_dir() and model_bin.is_file()
def _get_free_disk_space_bytes(path: Path) -> int:
"""Return free disk space in bytes for the filesystem containing path."""
path.mkdir(parents=True, exist_ok=True)
return shutil.disk_usage(path).free
def _load_model() -> WhisperModel:
"""Synchronous model load/download. Must run in a thread."""
MODELS_DIR.mkdir(parents=True, exist_ok=True)
VOICE_DIR.mkdir(parents=True, exist_ok=True)
model_path = MODELS_DIR / "base"
model_path = _get_model_path()
model_bin = model_path / "model.bin"
if not model_path.is_dir() or not model_bin.is_file():
_setup_proxy()
logger.info("Model not found locally, downloading to %s...", MODELS_DIR)
return WhisperModel(
"base",
logger.info(
"Whisper model '%s' not found locally. Starting download to %s...",
WHISPER_MODEL_SIZE,
MODELS_DIR,
)
start = time.monotonic()
model = WhisperModel(
WHISPER_MODEL_SIZE,
device="cpu",
cpu_threads=4,
compute_type="int8",
download_root=str(MODELS_DIR),
)
elapsed = time.monotonic() - start
logger.info(
"Whisper model '%s' downloaded and loaded in %.1fs",
WHISPER_MODEL_SIZE,
elapsed,
)
return model
logger.info("Loading model from %s...", model_path)
logger.info(
"Loading Whisper model '%s' from %s...",
WHISPER_MODEL_SIZE,
model_path,
)
return WhisperModel(
str(model_path),
device="cpu",
@@ -53,6 +88,44 @@ def _load_model() -> WhisperModel:
)
async def preload_model() -> None:
"""Download/load the Whisper model at bot startup if it is not already loaded."""
global _model
async with _model_lock:
if _model is not None:
logger.info("Whisper model is already loaded, skipping preload")
return
if _is_model_downloaded():
logger.info(
"Whisper model '%s' found locally at %s",
WHISPER_MODEL_SIZE,
_get_model_path(),
)
else:
free_space = _get_free_disk_space_bytes(MODELS_DIR)
required_space = WHISPER_MIN_FREE_SPACE_BYTES
logger.info(
"Free disk space: %.2f GB, required: %.2f GB",
free_space / (1024 ** 3),
required_space / (1024 ** 3),
)
if free_space < required_space:
raise RuntimeError(
f"Not enough disk space to download Whisper model '{WHISPER_MODEL_SIZE}'. "
f"Free: {free_space / (1024 ** 3):.2f} GB, "
f"required: {required_space / (1024 ** 3):.2f} GB"
)
logger.info(
"Whisper model '%s' will be downloaded at startup",
WHISPER_MODEL_SIZE,
)
loop = asyncio.get_running_loop()
_model = await loop.run_in_executor(None, _load_model)
logger.info("Whisper model '%s' is ready", WHISPER_MODEL_SIZE)
async def _get_model() -> WhisperModel:
"""Thread-safe lazy initializer for the Whisper model."""
global _model
@@ -61,7 +134,7 @@ async def _get_model() -> WhisperModel:
if _model is None:
loop = asyncio.get_running_loop()
_model = await loop.run_in_executor(None, _load_model)
logger.info("Whisper model loaded")
logger.info("Whisper model '%s' loaded on demand", WHISPER_MODEL_SIZE)
return _model
config.py
+5
View File
@@ -28,6 +28,11 @@ AI_BLOCK_DEFAULT_DURATION = 86400
AI_HEALTH_CHECK_ENABLED = os.getenv("AI_HEALTH_CHECK_ENABLED", "false").lower() == "true"
AI_HEALTH_CHECK_INTERVAL = int(os.getenv("AI_HEALTH_CHECK_INTERVAL", "600"))
# Whisper model settings.
# The base model requires ~2.5 GB for download; 5 GB free space is recommended.
WHISPER_MODEL_SIZE = os.getenv("WHISPER_MODEL_SIZE", "base")
WHISPER_MIN_FREE_SPACE_BYTES = int(os.getenv("WHISPER_MIN_FREE_SPACE_BYTES", str(5 * 1024 * 1024 * 1024)))
ACCESS_KEY = os.getenv("ACCESS_KEY", "").strip()
SECRET_KEY = os.getenv("SECRET_KEY", "").strip()
BUCKET_NAME = os.getenv("BUCKET_NAME", "").strip()
main.py
+9
View File
@@ -13,6 +13,7 @@ from bot.routers import moderation, layout, weather, ai, voice, yadisk, dialogue
from bot.utils.database import init_db, save_chat_user
from bot.utils.ai_client import start_model_health_check
from bot.utils.logging_config import setup_logging
from bot.utils.voice import preload_model
from bot.setup_commands import setup_bot_commands
from config import PROXY_ENABLED
@@ -75,6 +76,14 @@ async def main():
health_check_task.cancel()
return
try:
await preload_model()
logging.info("Модель Whisper готова к работе")
except Exception as exc:
logging.error("Не удалось загрузить модель Whisper: %s", exc)
health_check_task.cancel()
return
try:
await dp.start_polling(bot)
finally: