Dockerfile.eval

# Leoma EVAL SERVER image (CUDA).
#
# This is the GPU box: it downloads king + challenger weights from Hippius Hub,
# loads the pinned diffusers I2V pipeline, generates video on the held-out clips
# and scores each generation against the real continuation.
#
# It needs the [eval] extra (torch / diffusers / lpips / opencv / open_clip).
# The validator image (Dockerfile) deliberately does NOT.
FROM nvidia/cuda:12.4.1-cudnn-runtime-ubuntu22.04

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

# python3.12 + ffmpeg (ground-truth decode) + curl (healthcheck).
RUN apt-get update && apt-get install -y --no-install-recommends \
      software-properties-common \
 && add-apt-repository -y ppa:deadsnakes/ppa \
 && apt-get update && apt-get install -y --no-install-recommends \
      python3.12 python3.12-venv python3.12-dev \
      ffmpeg curl build-essential \
 && rm -rf /var/lib/apt/lists/*

RUN python3.12 -m venv /opt/venv
ENV PATH="/opt/venv/bin:${PATH}"

WORKDIR /app
RUN pip install --no-cache-dir uv

COPY pyproject.toml uv.lock README.md ./
COPY leoma ./leoma

# The [eval] extra is the whole point of this image. Frozen sync makes rebuilds use
# the exact tested CUDA/Diffusers/Transformers graph from uv.lock; an open-ended GPU
# stack can change generated frames even when Leoma's own source is unchanged.
RUN uv sync --active --frozen --extra eval --no-dev --no-editable --no-cache

# Fail at BUILD time if the eval stack can't import (torch/diffusers/chain.toml).
RUN python -c "import leoma.eval_server, leoma.eval.metrics, leoma.infra.chain_config as c; print('eval image OK, pipeline =', c.ARCH_PIPELINE)"

EXPOSE 9000
HEALTHCHECK --interval=30s --timeout=5s --start-period=60s --retries=3 \
  CMD curl -fsS -H "Authorization: Bearer ${LEOMA_EVAL_TOKEN}" \
      http://127.0.0.1:9000/health || exit 1

CMD ["leoma", "servers", "eval-server"]