Dockerfile.training

# Dockerfile.training — Dedicated GRPO Training HF Space
# ==========================================================
# Deploy this as a SEPARATE HF Space (hardware: A10G) to run
# train_grpo_hf.py on the Space's GPU without touching the
# production AEPO server Space.
#
# Create the training Space:
#   1. huggingface.co → New Space
#   2. Name: aepo-grpo-training  |  SDK: Docker  |  Hardware: A10G (24GB)
#   3. Add secrets: HF_TOKEN, HF_REPO
#   4. Upload this file as Dockerfile
#   5. The Space will clone the AEPO repo, train, push the LoRA adapter,
#      then serve a status page so you can check progress.

FROM nvidia/cuda:12.1.1-cudnn8-runtime-ubuntu22.04

ENV DEBIAN_FRONTEND=noninteractive \
    PYTHONUNBUFFERED=1 \
    PYTHONDONTWRITEBYTECODE=1 \
    PIP_NO_CACHE_DIR=1

RUN apt-get update && apt-get install -y --no-install-recommends \
        python3.11 python3.11-dev python3-pip git curl && \
    ln -sf /usr/bin/python3.11 /usr/bin/python && \
    apt-get clean && rm -rf /var/lib/apt/lists/*

WORKDIR /app

# ── Clone AEPO repo ────────────────────────────────────────────────────────────
ARG REPO_URL=https://github.com/umeshmaurya1301/autonomous-enterprise-payment-orchestrator.git
RUN git clone --depth 1 "${REPO_URL}" .

# ── Install AEPO runtime deps first (cached layer) ────────────────────────────
RUN pip install --upgrade pip && pip install -r requirements.txt

# ── Install GRPO training extras ──────────────────────────────────────────────
# Unsloth nightly wheels for CUDA 12.1 — matches the nvidia/cuda base image.
# xformers and flash-attn are optional; Unsloth falls back gracefully.
RUN pip install \
    "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git" \
    "trl>=0.15.0" \
    peft \
    accelerate \
    bitsandbytes \
    datasets \
    matplotlib \
    huggingface_hub

# ── Training entrypoint ───────────────────────────────────────────────────────
# Runs train_grpo_hf.py; after completion, starts a minimal HTTP server on
# port 7860 that serves the results directory so the judge / user can
# download grpo_reward_curve.png directly from the Space UI.
COPY Dockerfile.training.entrypoint.sh /entrypoint.sh
RUN chmod +x /entrypoint.sh

EXPOSE 7860
CMD ["/entrypoint.sh"]