跳转到内容

Mixed Precision Training

作者: Paulius Micikevicius, Sharan Narang, Jonah Alben, Greg Diamos, Erich Elsen, David Garcia, Boris Ginsburg, Michael Houston, Oleksii Kuchaiev, Ganesh Venkatesh, Hao Wu (2018)

arXiv: 1710.03740

领域

预训练

TLDR(中文)

系统论证了混合精度训练 recipe:FP16 存储与计算、FP32 主权重与累加、动态损失缩放防止梯度下溢。吞吐接近翻倍且精度无损,是大模型训练的效率基石;后来 BF16 免去损失缩放成为 LLM 标配,FP8 又在此基础上把吞吐再翻一倍。

TLDR (English)

Systematically validates mixed precision training: FP16 storage and compute with FP32 master weights and accumulators plus dynamic loss scaling to prevent gradient underflow. Nearly doubles throughput without quality loss; BF16 later removes the need for loss scaling and becomes the LLM default, and FP8 doubles throughput again.

出现在这些文章里

同被引用

这些论文与本文出现在同一篇文章中

相关论文

同一领域的其他论文