FROM python:3.12-slim

ENV PYTHONUNBUFFERED=1 \
    PYTHONDONTWRITEBYTECODE=1 \
    PIP_NO_CACHE_DIR=1

WORKDIR /app

# torch first, from the CPU-only index. `--extra-index-url` would not be enough:
# PyPI publishes the same version number as a CUDA build, and pip is free to
# resolve to it -- which is 2.5GB of CUDA runtime this Space will never call.
RUN pip install --no-cache-dir --index-url https://download.pytorch.org/whl/cpu torch==2.13.0
COPY requirements.txt .
RUN pip install --no-cache-dir -r requirements.txt

# Bake the weights into the image: a Space that pulls a 409MB checkpoint on
# every cold start answers its first message a minute late, and the trimmed
# file is 136MB.
COPY prepare_weights.py .
RUN python prepare_weights.py /app/weights

COPY booper ./booper
COPY static ./static
COPY app.py .

EXPOSE 7860
# One worker: the model is a shared in-process object and 2 vCPUs are already
# saturated by a single decode, so a second worker would only double the RAM.
CMD ["uvicorn", "app:app", "--host", "0.0.0.0", "--port", "7860", "--workers", "1", "--no-access-log"]