1 paper · 1 filter
Shwai He, Tao Ge, Guoheng Sun +3
Traditional transformer models often allocate a fixed amount of computational resources to every input token, leading to inefficient and unnecessary computation. To address this, t…