docker-compose.yml

version: "3.8"

services:
  # The GPU-accelerated vLLM Engine (needs an NVIDIA GPU; omit on CPU-only boxes)
  vllm-engine:
    image: vllm/vllm-openai:latest
    container_name: math2code_vllm
    command: >
      --model AI-MO/NuminaMath-7B-TIR
      --enable-lora
    ports:
      - "8080:8000"
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]
    volumes:
      - ~/.cache/huggingface:/root/.cache/huggingface
      - ./outputs/aimo_lora/final:/adapters/math2code_lora

  # FastAPI proxy: vLLM completion -> extract code -> sandboxed execution
  api:
    build: .
    container_name: math2code_api
    command: uvicorn math2code.serve.api:app --host 0.0.0.0 --port 8000
    ports:
      - "8000:8000"
    depends_on:
      - vllm-engine
    environment:
      - VLLM_API_BASE=http://vllm-engine:8000/v1
      - VLLM_MODEL_NAME=AI-MO/NuminaMath-7B-TIR

  # Gradio UI
  frontend:
    build: .
    container_name: math2code_frontend
    command: python -m math2code.serve.app
    ports:
      - "8501:8501"
    depends_on:
      - api
    environment:
      - API_URL=http://api:8000/generate