Linearly Decomposing and Recomposing Vision Transformers for Diverse-Scale Models | Read Paper on Bytez