VLM efficiency

V-CoLA: Compressing Vision Tokens Using Linear Attention

V-CoLA compresses vision tokens in linear-attention VLMs without training, keeping accuracy and speeding prefi…