AIM: Adaptive Inference of Multi-Modal LLMs via Token Merging and Pruning