Dockerfile

# HF Spaces GPU container for GRPO training.
#
# Build context = repo root (fake_gang_env/). The Space runs the env server +
# launches a training phase. The phase is selected via the PHASE env var.
#
# Default workflow on the Space:
#   1) start the env server in the background (uvicorn, port 8000)
#   2) wait for /health
#   3) run `python -m training.train_grpo --phase $PHASE ...`
#
# To run locally:  docker build -t fakegang-train -f training/Dockerfile .
#                  docker run --gpus all -e PHASE=phase1 fakegang-train

FROM nvidia/cuda:12.4.1-cudnn-runtime-ubuntu22.04

ENV DEBIAN_FRONTEND=noninteractive \
    PIP_NO_CACHE_DIR=1 \
    PYTHONUNBUFFERED=1 \
    HF_HUB_ENABLE_HF_TRANSFER=1 \
    PHASE=phase1 \
    MODEL=Qwen/Qwen2.5-1.5B-Instruct \
    PLATFORM=Instagram \
    ENV_BASE_URL=https://pandago-graphstrike-model-training.hf.space

RUN apt-get update && apt-get install -y --no-install-recommends \
        python3 python3-pip python3-venv git curl ca-certificates && \
    rm -rf /var/lib/apt/lists/*
RUN ln -sf /usr/bin/python3 /usr/bin/python

WORKDIR /app

# Install training deps. Build arg CACHE_BUST forces this layer to rebuild.
# Bump the default (or pass --build-arg CACHE_BUST=$(date +%s)) when pinned
# versions change but Spaces is reusing the cached pip layer.
ARG CACHE_BUST=2026-04-25-c
RUN echo "cache-bust: $CACHE_BUST"
COPY training/requirements.txt /tmp/training-requirements.txt
RUN pip install --upgrade pip && \
    pip install --no-cache-dir --upgrade --force-reinstall torch==2.6.0 && \
    pip install --no-cache-dir --upgrade --force-reinstall -r /tmp/training-requirements.txt && \
    python3 -c "import torch; from torch.distributed.fsdp import FSDPModule; print('torch', torch.__version__, 'FSDPModule OK')" && \
    python3 -c "import trl; print('trl', trl.__version__)"

# Copy the rest of the repo (client.py + eval-models/_round2_runner.py + training/).
COPY . /app

EXPOSE 7860

CMD ["bash", "training/entrypoint.sh"]