Dockerfile

# Hugging Face Docker Space — optional GPU path (prebuilt wheel, no source compile).
# Default public Space uses Gradio SDK; see README.md sdk: gradio.
ARG CUDA_IMAGE=12.4.1-runtime-ubuntu22.04
FROM nvidia/cuda:${CUDA_IMAGE}

ENV DEBIAN_FRONTEND=noninteractive \
    PYTHONUNBUFFERED=1 \
    SYSTEM=spaces \
    HF_HUB_ENABLE_HF_TRANSFER=1

RUN apt-get update && apt-get install -y --no-install-recommends \
    python3 \
    python3-pip \
    && rm -rf /var/lib/apt/lists/*

WORKDIR /app
COPY . /app

RUN grep -v '^llama-cpp-python' requirements.txt > /tmp/requirements-core.txt \
    && pip3 install --no-cache-dir -r /tmp/requirements-core.txt

# Prebuilt CUDA wheel — avoids OOM from compiling llama-cpp-python at build time.
# >= 0.3.25 required for Nemotron 3 Nano (nemotron_h) GGUF support.
RUN pip3 install --no-cache-dir "llama-cpp-python==0.3.29" \
    --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cu124

RUN chmod +x /app/scripts/space_entrypoint.sh

ENV PMS_INFERENCE_BACKEND=llamacpp \
    PMS_LLAMACPP_MODE=embedded \
    PMS_USE_FINETUNED=true \
    PMS_FINETUNED_MODEL=GuusBouwensNL/plane-mode-nemotron-4b-study-coach \
    PMS_LORA_GGUF_MODEL=GuusBouwensNL/plane-mode-study-coach-lora-gguf \
    PMS_GGUF_DIR=/data/models/nemotron \
    PMS_LLAMACPP_N_GPU_LAYERS=-1 \
    GRADIO_SERVER_NAME=0.0.0.0

RUN useradd -m -u 1000 user && chown -R user:user /app
USER user
ENV HOME=/home/user PATH=/home/user/.local/bin:${PATH}

EXPOSE 7860
CMD ["/app/scripts/space_entrypoint.sh"]