FROM python:3.11-slim

WORKDIR /app

# Install CPU-only torch first (saves ~1.8GB vs CUDA build)
RUN pip install --no-cache-dir torch --index-url https://download.pytorch.org/whl/cpu

COPY requirements.txt .
RUN pip install --no-cache-dir -r requirements.txt

COPY app.py .

# Pre-download commonly used TTS models so first requests are fast.
# Individual failures are non-fatal.
RUN python -c "\
from transformers import VitsModel, AutoTokenizer; \
models = [ \
    'facebook/mms-tts-yor', 'facebook/mms-tts-swh', 'facebook/mms-tts-hau', \
    'facebook/mms-tts-pcm', 'facebook/mms-tts-aka', 'facebook/mms-tts-lug', \
    'facebook/mms-tts-amh', 'facebook/mms-tts-som', 'facebook/mms-tts-sna', \
    'khof312/mms-tts-lin', 'facebook/mms-tts-ara', \
]; \
ok = 0; \
for m in models: \
    try: \
        VitsModel.from_pretrained(m); AutoTokenizer.from_pretrained(m); \
        print(f'  OK {m}'); ok += 1 \
    except Exception as e: \
        print(f'  SKIP {m}: {e}') \
; \
print(f'{ok}/{len(models)} models cached') \
"

EXPOSE 7860

CMD ["uvicorn", "app:app", "--host", "0.0.0.0", "--port", "7860"]