CASP: Compression of Large Multimodal Models Based on Attention Sparsity | Read Paper on Bytez