services:
  # 1. Hardware Ops & MFU Smoketest Profiler
  smoketest:
    build:
      context: .
      dockerfile: Dockerfile
    image: yoshida:latest
    container_name: yoshida_smoketest
    ipc: host
    shm_size: '16gb'
    command: python3 smoketest.py --config configs/smoketest.yaml
    volumes:
      - ~/.cache/huggingface:/root/.cache/huggingface
      - ./saved_models:/workspace/yoshida/saved_models
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]

  # 2. DPO Safety Alignment Training
  train-dpo:
    build:
      context: .
      dockerfile: Dockerfile
    image: yoshida:latest
    container_name: yoshida_train_dpo
    ipc: host
    shm_size: '16gb'
    command: python3 train.py --config configs/dpo_safety.yaml
    volumes:
      - ~/.cache/huggingface:/root/.cache/huggingface
      - ./saved_models:/workspace/yoshida/saved_models
    environment:
      - WANDB_API_KEY=${WANDB_API_KEY:-}
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]

  # 3. GRPO Safety Alignment Training (DeepSeek R1 / Safety style)
  train-grpo:
    build:
      context: .
      dockerfile: Dockerfile
    image: yoshida:latest
    container_name: yoshida_train_grpo
    ipc: host
    shm_size: '16gb'
    command: python3 train.py --config configs/grpo_safety.yaml
    volumes:
      - ~/.cache/huggingface:/root/.cache/huggingface
      - ./saved_models:/workspace/yoshida/saved_models
    environment:
      - WANDB_API_KEY=${WANDB_API_KEY:-}
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]

  # 4. Multi-GPU FSDP (ZeRO-3) Distributed Training
  train-fsdp:
    build:
      context: .
      dockerfile: Dockerfile
    image: yoshida:latest
    container_name: yoshida_train_fsdp
    ipc: host
    shm_size: '32gb'
    command: torchrun --nproc_per_node=4 train.py --config configs/fsdp_grpo_safety.yaml
    volumes:
      - ~/.cache/huggingface:/root/.cache/huggingface
      - ./saved_models:/workspace/yoshida/saved_models
    environment:
      - WANDB_API_KEY=${WANDB_API_KEY:-}
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]

  # 5. Interactive Development Shell
  dev:
    build:
      context: .
      dockerfile: Dockerfile
    image: yoshida:latest
    container_name: yoshida_dev
    ipc: host
    shm_size: '16gb'
    stdin_open: true
    tty: true
    volumes:
      - .:/workspace/yoshida
      - ~/.cache/huggingface:/root/.cache/huggingface
      - ./saved_models:/workspace/yoshida/saved_models
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]