LLaVA-Phi: Efficient Multi-Modal Assistant with Small Language Model