Dockerfile.train

# Training-only image for X-VLA fine-tuning on H100
# Does NOT include IsaacSim — evaluation runs locally
FROM pytorch/pytorch:2.5.1-cuda12.4-cudnn9-devel

ENV DEBIAN_FRONTEND=noninteractive
ENV PYTHONUNBUFFERED=1
ENV HF_HOME=/workspace/hf_cache

WORKDIR /workspace

# System deps
RUN apt-get update && apt-get install -y \
    git curl wget rsync \
    libglib2.0-0 libsm6 libxrender1 libxext6 \
    && rm -rf /var/lib/apt/lists/*

# Install lerobot with xvla extras
# Pin transformers: >=4.47 required for Florence2LanguageConfig + huggingface-hub 1.x
RUN pip install --no-cache-dir \
    "lerobot[xvla]" \
    "transformers>=4.47.0,<5.0.0" \
    draccus \
    huggingface_hub \
    wandb

# Copy project code (datasets/outputs/Assets mounted as volumes at runtime)
COPY configs/ /workspace/configs/
COPY scripts/ /workspace/scripts/
COPY source/ /workspace/source/
COPY pyproject.toml /workspace/

CMD ["bash"]