Towards Efficient Large Multimodal Model Serving