1
Rethinking Shrinkage Bias in LLM FP4 Pretraining: Geometric Origin, Systemic Impact, and UFP4 Recipe
直击FP4低精度训练痛点，首次从几何角度解释收缩偏差的根源，并提出UFP4方案系统性解决该问题。
arXiv:2606.20381v1 Announce Type: new Abstract: FP4 training promises substantial reductions in memory and computation cost for LLM pretraining, yet c…