Dockerfile

FROM python:3.12-slim

# Non-root user (HF Spaces convention; writable HOME for the HF cache the model
# is pulled into at boot).
# GLINER_MOCK=0: the deployed Space runs live inference (the local dev kit
# defaults to mock; the env var flips only this image).
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 \
    GLINER_MOCK=0

WORKDIR /home/user/app

# git needed to pip-install gliner2 from GitHub branch
USER root
RUN apt-get update && apt-get install -y --no-install-recommends git && rm -rf /var/lib/apt/lists/*
USER user

# 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 protobuf
# GLiNER2 from main branch (required for boundary architecture support)
RUN pip install --no-cache-dir --user "gliner2[local] @ git+https://github.com/fastino-ai/GLiNER2.git@main"

COPY --chown=user . .

# Pre-download model at build time to speed up cold starts. The model repo is
# public on the HF hub, so no token is needed.
RUN python -c "from huggingface_hub import snapshot_download; snapshot_download('fastino/gliner2.5-multi-v1')" || echo "model pre-download skipped"

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"]