Dockerfile.gpu

# VideoAnnotator Production Docker Image - GPU Version
# This image does NOT include models/weights - they download automatically on first use
#
# Usage:
#   docker build -f Dockerfile.gpu -t videoannotator:gpu .
#   docker run --gpus all --rm -p 8000:8000 -v ${PWD}/data:/app/data videoannotator:gpu

FROM nvidia/cuda:12.6.0-runtime-ubuntu24.04

# Use bash with pipefail so RUN commands that use a pipe fail when any stage does
SHELL ["/bin/bash","-o","pipefail","-lc"]

# Prevent interactive prompts during package installation
ARG DEBIAN_FRONTEND=noninteractive

# Install base packages and locales in a single RUN to reduce layers and avoid
# pulling recommended packages unnecessarily (DL3015, DL3059)
RUN apt-get update && apt-get install -y --no-install-recommends \
    curl python3 python3-venv python3-dev git git-lfs ffmpeg \
    libgl1-mesa-dri libglib2.0-0 libsm6 libxext6 libxrender1 libgomp1 locales \
    && locale-gen en_US.UTF-8 \
    && update-locale LANG=en_US.UTF-8 LC_ALL=en_US.UTF-8 \
    && rm -rf /var/lib/apt/lists/* \
    && git lfs install

# Export UTF-8 locale for all processes
ENV LANG=en_US.UTF-8
ENV LC_ALL=en_US.UTF-8
ENV UV_LINK_MODE=copy

# uv package manager
RUN curl -LsSf https://astral.sh/uv/install.sh | sh
ENV PATH="/root/.local/bin:${PATH}"

WORKDIR /app

# Build args to allow dev builds to skip network-heavy steps (defaults: true)
# For production image builds set these to "false" to perform installs in image.
ARG SKIP_IMAGE_UV_SYNC=true
ARG SKIP_TORCH_INSTALL=true

# Copy project definition and lock file first to leverage Docker cache
COPY pyproject.toml uv.lock ./

# Install dependencies (including torch from configured index)
# These steps can be skipped at build time for dev containers by setting
# build args SKIP_IMAGE_UV_SYNC=true. Post-create commands will run `uv sync`
# inside the running container where network is usually available.
RUN if [ "${SKIP_IMAGE_UV_SYNC}" != "true" ]; then uv sync --frozen --no-install-project --no-editable; else echo "[BUILD] Skipping uv sync at image build (SKIP_IMAGE_UV_SYNC=true)"; fi

# Copy the rest of the source code (explicitly exclude models/weights)
COPY api_server.py ./
COPY src/ ./src/
COPY configs/ ./configs/
COPY scripts/ ./scripts/

# Install hadolint into a system location so pre-commit can find it.
RUN HADOLINT_DEST_DIR=/usr/local/bin bash scripts/install_hadolint.sh

# Verify GPU access (optional)
# NOTE: Some builds may intentionally skip Torch installation/sync to speed up
# devcontainer image builds. In that case, skip the Torch import check.
RUN if [ "${SKIP_TORCH_INSTALL}" = "true" ]; then \
        echo "[BUILD] Skipping Torch verification (SKIP_TORCH_INSTALL=true)"; \
    else \
        uv run python3 -c "import torch; print(f'[GPU BUILD] CUDA available: {torch.cuda.is_available()}'); print(f'[GPU BUILD] PyTorch version: {torch.__version__}'); print('[GPU BUILD] Production image ready - models will download on first use')"; \
    fi

# Set environment for production
ENV PYTHONUNBUFFERED=1
ENV CUDA_VISIBLE_DEVICES=0

# Create directories for mounted volumes
RUN mkdir -p /app/data /app/output /app/logs

EXPOSE 18011

CMD ["uv", "run", "python3", "api_server.py", "--log-level", "info", "--port", "18011"]