Parameter-efficient Tuning of Large-scale Multimodal Foundation Model