services:
  llamafactory:
    build:
      dockerfile: ./docker/docker-xpu/Dockerfile
      context: ../..
      args:
        PIP_INDEX: https://pypi.org/simple
        # The DLE base image determines which Intel GPU runtime (libze-intel-gpu)
        # is shipped. The runtime version must be >= the host driver version,
        # otherwise torch.xpu.device_count() returns 0:
        #   DLE 2025.3 → libze-intel-gpu 25.18.x  (works on driver ≤26.09)
        #   DLE 2026.1 → libze-intel-gpu 26.18.x+ (works on driver 26.18/26.22)
        # Verify host driver: dpkg -l libze-intel-gpu1 | grep -oP '\d+\.\d+\.\d+'
        BASE_IMAGE: intel/deep-learning-essentials:2026.1.0-devel-ubuntu24.04
    container_name: llamafactory
    image: llamafactory:xpu
    ports:
      - "7860:7860"
      - "8000:8000"
    ipc: host
    tty: true
    # shm_size: "16gb"  # ipc: host is set
    stdin_open: true
    command: bash
    devices:
      # Intel GPU character devices (renderD* and card* under /dev/dri).
      # renderD* nodes are owned by group `render`; card* by group `video`.
      - /dev/dri:/dev/dri
    volumes:
      # by-path symlinks are required for oneCCL's ze_fd_manager (2-GPU IPC).
      # `devices:` alone does not carry sub-directories — this volume does.
      - /dev/dri/by-path:/dev/dri/by-path
      - ~/.cache/huggingface:/root/.cache/huggingface
    group_add:
      # Grant access to the Intel GPU device nodes.
      # renderD* nodes are owned by group `render`; card* nodes by `video`.
      # Use numeric GIDs here — Docker compose resolves group names against
      # the CONTAINER's /etc/group (not the host), and the DLE base image
      # does not carry a `render` group entry. Standard Linux GIDs:
      # If your host uses different GIDs, run `getent group render video` and
      # update these values accordingly.
      # example output.
      # getent group render video
      # render:x:992:root
      # video:x:44:support
      - "992"   # render — owns /dev/dri/renderD* nodes
      - "44"    # video  — owns /dev/dri/card* nodes
    restart: unless-stopped