近日，Radical Numerics 与蒙特利尔大学 Yoshua Bengio 团队找了一个新思路，为 LLM 的效率进化提供了一个极具启发性的工程视角。该团队通过将线性递归重新定义为硬件对齐的矩阵运算，提出了一套能够相当完美契合 GPU 内存层级的算法框架。

Let $t_0, t_1, t_2, \cdots$ be a sequence of elements of a field $F$. We give a continued fraction algorithm for $t_0x + t_1x^2 + t_2x^3 + \cdots$. If our sequence ...

This is a preview. Log in through your library . Journal Information Mathematica Scandinavica has been published regularly since 1953. Mathematica Scandinavica is a non-profit company, which is ...