# Benchmark evaluation services for LeRobot.
#
# Each service builds Dockerfile.benchmark with the matching BENCHMARK arg,
# giving an isolated environment per benchmark suite.
#
# Usage:
#   Build all:        docker compose -f docker/docker-compose.benchmark.yml build
#   Build one:        docker compose -f docker/docker-compose.benchmark.yml build libero
#   Run smoke tests:  docker compose -f docker/docker-compose.benchmark.yml run --rm libero bash docker/smoke_test_benchmark.sh
#   Interactive:      docker compose -f docker/docker-compose.benchmark.yml run --rm libero

x-benchmark-base: &benchmark-base
  build:
    context: .
    dockerfile: docker/Dockerfile.benchmark
  image: lerobot-benchmark-${BENCHMARK:-libero}
  environment:
    - MUJOCO_GL=egl
    - DEVICE=cuda
  deploy:
    resources:
      reservations:
        devices:
          - driver: nvidia
            count: 1
            capabilities: [gpu]
  volumes:
    # Mount HuggingFace cache from host so model weights are not re-downloaded each run.
    - ${HF_HOME:-~/.cache/huggingface}:/home/user_lerobot/.cache/huggingface
  stdin_open: true
  tty: true

services:
  libero:
    <<: *benchmark-base
    image: lerobot-benchmark-libero
    build:
      context: .
      dockerfile: docker/Dockerfile.benchmark
      args:
        BENCHMARK: libero

  libero_plus:
    <<: *benchmark-base
    image: lerobot-benchmark-libero-plus
    build:
      context: .
      dockerfile: docker/Dockerfile.benchmark
      args:
        BENCHMARK: libero_plus

  robomme:
    <<: *benchmark-base
    image: lerobot-benchmark-robomme
    build:
      context: .
      dockerfile: docker/Dockerfile.benchmark
      args:
        BENCHMARK: robomme

  robocasa:
    <<: *benchmark-base
    image: lerobot-benchmark-robocasa
    build:
      context: .
      dockerfile: docker/Dockerfile.benchmark
      args:
        BENCHMARK: robocasa