CurveFP：用于语言模型的封闭乘积有理基数对数数据类型
原标题：CurveFP: Rational-Radix Logarithmic Datatypes with Closed Products for Language Models
AI 摘要
CurveFP 是一种新的低精度数据类型家族，通过交错对数曲线和紧凑块缩放实现封闭乘积，使乘法变为精确的符号异或和整数索引更新。CurveFP8 用于训练，CurveFP7 用于部署，在多个 7B-9B 模型上以更少的位达到或超过 FP8 的困惑度，并在 128.3M 参数预训练中实现 FP8 级质量，同时降低格式引起的惩罚。
正文节选
CurveFP: Rational-Radix Logarithmic Datatypes with Closed Products for Language Models Abstract Low-precision datatypes reduce language-model cost, but most formats optimize scalar fidelity while leaving the arithmetic induced by their products unchanged. We introduce CurveFP, a closed-product codebook family that distributes quantized magnitudes across interleaved logarithmic curves under compact block scales. A rational radix tunes dynamic range against local resolution, while uniform curve in