docker_entry.py

# Hugging Face Space (Docker) — GenSearcher + FireRed
# Requires GPU. For multi-GPU full-local mode, set START_VLLM_*=1 and CUDA device envs in README.

FROM pytorch/pytorch:2.5.1-cuda12.4-cudnn9-runtime

ENV DEBIAN_FRONTEND=noninteractive
RUN apt-get update && apt-get install -y --no-install-recommends \
    curl \
    git \
    && rm -rf /var/lib/apt/lists/*

WORKDIR /app

COPY vendor/rllm /app/vendor/rllm
COPY requirements.txt /app/requirements.txt
COPY app.py space_gen.py space_health.py /app/
COPY services /app/services
COPY scripts /app/scripts
COPY docker_entry.py /app/docker_entry.py
# Strip BOM + all CR so bash never sees \r (avoids "exec format error" / stray $'\r' errors).
RUN python3 -c "import pathlib; p=pathlib.Path('/app/scripts/entrypoint.sh'); b=p.read_bytes(); b=b.lstrip(b'\\xef\\xbb\\xbf'); b=b.replace(b'\\r\\n', b'\\n').replace(b'\\r', b''); p.write_bytes(b)" \
    && chmod +x /app/scripts/entrypoint.sh

ENV PYTHONPATH=/app/vendor/rllm
ENV GRADIO_SERVER_PORT=7860
# HF Spaces / minimal images often have uid 1000 with no /etc/passwd entry; PyTorch Inductor calls
# getpass.getuser() and crashes with KeyError. USER/LOGNAME short-circuit getuser(); cache dirs avoid $HOME issues.
ENV USER=huggingface
ENV LOGNAME=huggingface
ENV TORCHINDUCTOR_CACHE_DIR=/tmp/torch_inductor_cache
ENV TRITON_CACHE_DIR=/tmp/triton_cache

RUN pip install --no-cache-dir --upgrade pip setuptools wheel \
    && pip install --no-cache-dir -e /app/vendor/rllm \
    && pip install --no-cache-dir -r /app/requirements.txt

# Optional: local vLLM inside the image (large). Disable with build-arg if you only use external APIs.
ARG INSTALL_VLLM=1
RUN if [ "$INSTALL_VLLM" = "1" ]; then pip install --no-cache-dir "vllm>=0.6.3"; fi

EXPOSE 7860

# Python as PID 1: kernel never execve's entrypoint.sh directly (fixes HF "exec format error" on CRLF/BOM).
CMD ["python", "/app/docker_entry.py"]