Skipping Computations in Multimodal LLMs