Dockerfile.gh200

# Base: NGC vLLM container for aarch64 with CUDA 13.1, Python 3.12
FROM nvcr.io/nvidia/vllm:25.12.post1-py3

ARG DEBIAN_FRONTEND=noninteractive

# Unset ROCR_VISIBLE_DEVICES to prevent conflict with CUDA_VISIBLE_DEVICES
# (cluster sets both, but verl requires only one)
ENV ROCR_VISIBLE_DEVICES=""
ENV NCCL_NET=Socket

# Install system dependencies
RUN apt-get update && apt-get install -y --no-install-recommends \
    git \
    wget \
    curl \
    build-essential \
    libsndfile1 \
    libgl1 \
    libglib2.0-0 \
    && rm -rf /var/lib/apt/lists/*

WORKDIR /app

# Copy requirements file first (for layer caching)
COPY requirements-gh200.txt /app/requirements-gh200.txt

# Install Python dependencies (excluding NGC pre-installed packages)
RUN pip install --no-cache-dir -r requirements-gh200.txt

# Copy project files
COPY . /app

# Install verl/SDPO in editable mode (no-deps to avoid reinstalling dependencies)
RUN pip install -e . --no-deps

# Default command
CMD ["/bin/bash"]