docker-compose.yml

version: '3.8'

services:
  latentroute-train:
    image: latentroute:latest
    container_name: latentroute-train
    build:
      context: .
      dockerfile: Dockerfile
    
    # GPU support
    runtime: nvidia
    environment:
      - NVIDIA_VISIBLE_DEVICES=all
      - NVIDIA_DRIVER_CAPABILITIES=compute,utility
    
    # Load environment variables
    env_file:
      - .env.docker
    
    # Volume mounts
    volumes:
      # Model checkpoints and final model
      - ./models:/models
      - /workspace/models:/models
      
      # Training data (wiki_corpus.jsonl, tokenizer vocab)
      - ./data:/data
      - /workspace/data:/data
      
      # Cache directories (large - HF datasets, torch)
      - cache-huggingface:/cache/huggingface
      - cache-torch:/cache/torch
      
      # Logs
      - ./logs:/workspace/logs
      - /workspace/logs:/workspace/logs
      
      # Optional: Ray tune results
      - ray-results:/workspace/ray_results
    
    # Expose Ray Tune dashboard (if using distributed training)
    ports:
      - "8265:8265"  # Ray Tune dashboard
      - "8888:8888"  # Jupyter (optional)
    
    # Resource limits (adjust based on your hardware)
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1          # Number of GPUs to allocate
              capabilities: [gpu]
        limits:
          memory: 120G          # Max memory allocation
    
    # Keep container running
    stdin_open: true
    tty: true
    
    # Restart policy
    restart: unless-stopped
    
    # Override entrypoint for interactive shell if needed
    # command: /bin/bash

  # Optional: Separate data preparation service (CPU only)
  latentroute-prepare:
    image: latentroute:latest
    build:
      context: .
      dockerfile: Dockerfile
    
    container_name: latentroute-prepare
    
    env_file:
      - .env.docker
    
    volumes:
      - ./data:/data
      - /workspace/data:/data
      - cache-huggingface:/cache/huggingface
    
    # No GPU needed for data preparation
    command: /bin/bash
    
    stdin_open: true
    tty: true
    
    profiles:
      - data-prep

# Named volumes for persistent storage across container restarts
volumes:
  cache-huggingface:
    driver: local
  cache-torch:
    driver: local
  ray-results:
    driver: local

networks:
  default:
    name: latentroute-network
    driver: bridge