services:
  tabbyapi:
    # Uncomment this to build a docker image from source
    #build:
    #  context: ..
    #  dockerfile: ./docker/Dockerfile
    #  # For CUDA 13.x builds, use ./docker/Dockerfile.cu13
    #  # For AMD GPUs, see the tabbyapi-rocm service below
    #  # Uncomment to include the optional embeddings stack (infinity-emb)
    #  #args:
    #  #  EXTRAS: 1

    # Comment this to build a docker image from source
    image: ghcr.io/theroyallab/tabbyapi:latest
    gpus: all
    ports:
      - "5000:5000"
    # ExLlamaV3 keeps tensor-parallel and CPU MoE offload buffers in /dev/shm (and
    # NCCL uses it as a transport). Docker's default of 64 MiB is far too small.
    # This is an upper bound on a tmpfs, not a reservation.
    shm_size: "8g"
    ulimits:
      memlock: -1
      nofile: 1048576
    healthcheck:
      test: ["CMD", "curl", "-f", "http://127.0.0.1:5000/health"]
      interval: 30s
      timeout: 10s
      retries: 3
    environment:
      - NAME=TabbyAPI
      - NVIDIA_VISIBLE_DEVICES=all
    volumes:
      - ./models:/app/models                                # Change me
      # - /path/to/config.yml:/app/config.yml               # Change me
      # - /path/to/api_tokens.yml:/app/api_tokens.yml       # Change me

  # AMD GPUs. Start with: docker compose --profile rocm up
  # The image is built per GPU target (see docker/Dockerfile.rocm); the host needs
  # only the amdgpu kernel driver, the ROCm runtime is inside the image.
  tabbyapi-rocm:
    profiles: ["rocm"]
    # Uncomment this to build a docker image from source
    #build:
    #  context: ..
    #  dockerfile: ./docker/Dockerfile.rocm
    #  args:
    #    ROCM_TARGETS: gfx1100          # e.g. gfx1201, gfx1030, gfx1151; several: "gfx1100 gfx1201"
    #    # EXTRAS: 1

    # Comment this to build a docker image from source
    image: ghcr.io/theroyallab/tabbyapi:latest-rocm
    devices:
      - /dev/kfd
      - /dev/dri
    # The container user needs the host's video and render groups to open the
    # GPU devices. Images don't define a render group, so give its numeric id:
    # RENDER_GID=$(getent group render | cut -d: -f3) docker compose --profile rocm up
    group_add:
      - video
      - "${RENDER_GID:-44}"
    ports:
      - "5000:5000"
    shm_size: "8g"
    ulimits:
      memlock: -1
      nofile: 1048576
    healthcheck:
      test: ["CMD", "curl", "-f", "http://127.0.0.1:5000/health"]
      interval: 30s
      timeout: 10s
      retries: 3
    environment:
      - NAME=TabbyAPI
    volumes:
      - ./models:/app/models                                # Change me
      # - /path/to/config.yml:/app/config.yml               # Change me
      # - /path/to/api_tokens.yml:/app/api_tokens.yml       # Change me