BriefGPT.xyz
Oct, 2024
Syno:神经算子的结构合成
Syno: Structured Synthesis for Neural Operators
HTML
PDF
Yongqi Zhuo, Zhengyuan Su, Chenggang Zhao, Mingyu Gao
TL;DR
本研究针对现有神经网络架构和张量编译器在设计新算子上的局限性,提出了一种自动高效发现新神经算子的框架Syno。Syno通过细粒度原语和引导合成流程,显著提升了算子的速度和准确性,测试结果显示,与经过NAS优化的模型相比,速度提升平均达到2.06倍,且准确率损失小于1%。
Abstract
The desires for better prediction accuracy and higher execution performance in neural networks never end.
Neural Architecture Search
(NAS) and
Tensor Compilers
are two popular techniques to optimize these two goa
→