Dockerfile

FROM nvcr.io/nvidia/pytorch:25.10-py3

WORKDIR /app

RUN pip install --no-cache-dir \
    fastapi \
    "uvicorn[standard]" \
    transformers \
    accelerate \
    huggingface_hub

# Pre-download model at build time so startup is instant
RUN python3 -c "from transformers import AutoTokenizer, AutoModelForCausalLM; AutoTokenizer.from_pretrained('OpenmindAGI/functiongemma-finetuned-g1-multilingual'); AutoModelForCausalLM.from_pretrained('OpenmindAGI/functiongemma-finetuned-g1-multilingual')"
COPY src/functiongemma/server.py .

# Jetson Thor inference optimizations
ENV CUDA_MODULE_LOADING=LAZY
ENV PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
ENV OMP_NUM_THREADS=4
ENV TOKENIZERS_PARALLELISM=false

EXPOSE 8200

HEALTHCHECK --interval=30s --timeout=10s --start-period=120s --retries=3 \
    CMD python3 -c "import urllib.request; urllib.request.urlopen('http://localhost:8200/health')" || exit 1

CMD ["uvicorn", "server:app", "--host", "0.0.0.0", "--port", "8200", "--workers", "1"]