Dockerfile

FROM vllm/vllm-openai:latest

ENV DEBIAN_FRONTEND=noninteractive
ENV PYTHONUNBUFFERED=1

ENV MODEL_NAME=NM-dev/NuExtract3.4_4B-RL-400
ENV OPENAI_API_BASE=http://127.0.0.1:8000/v1
ENV OPENAI_API_KEY=EMPTY

ENV GRADIO_SERVER_NAME=0.0.0.0
ENV GRADIO_SERVER_PORT=7860

ENV VLLM_PORT=8000
ENV MAX_MODEL_LEN=8192
ENV GPU_MEMORY_UTILIZATION=0.90

ENV NUEXTRACT_MAX_TOKENS=5000
ENV NUEXTRACT_ASSETS_DIR=/home/user/app/assets

ENV HF_HOME=/data/.cache/huggingface
ENV TRANSFORMERS_CACHE=/data/.cache/huggingface
ENV VLLM_CACHE_ROOT=/data/.cache/vllm

RUN pip install --no-cache-dir \
    gradio \
    openai \
    pillow

RUN useradd -m -u 1000 user || true

WORKDIR /home/user/app

COPY --chown=user:user app.py /home/user/app/app.py
COPY --chown=user:user start.sh /home/user/app/start.sh
COPY --chown=user:user assets /home/user/app/assets

RUN chown -R user:user /home/user/app && \
    chmod +x /home/user/app/start.sh

USER user

EXPOSE 7860

# Important: override vllm/vllm-openai's default entrypoint.
ENTRYPOINT []

CMD ["/bin/bash", "/home/user/app/start.sh"]