FROM python:3.12-slim

# Non-root user (HF Spaces convention; writable HOME for the HF cache the model
# is pulled into at boot).
RUN useradd -m -u 1000 user
USER user
ENV HOME=/home/user \
    PATH=/home/user/.local/bin:$PATH \
    HF_HOME=/home/user/.cache/huggingface \
    PYTHONUNBUFFERED=1

WORKDIR /home/user/app

# CPU-only torch (default PyPI wheel is CUDA and huge). Only needed for the real
# model; the mock path works on flask alone.
RUN pip install --no-cache-dir --user torch --index-url https://download.pytorch.org/whl/cpu
RUN pip install --no-cache-dir --user flask gunicorn

COPY --chown=user . .

EXPOSE 7860

# One worker (single shared model), threads for static/infer connections.
# No --preload: avoids the torch/OpenMP fork-after-threads deadlock.
CMD ["gunicorn", "-w", "1", "--threads", "8", "-k", "gthread", "-b", "0.0.0.0:7860", "--timeout", "0", "app:app"]