Mingchuan Ma

Full-Stack FP4: Stable LLM Pretraining with Quantized Projections, Optimizers, and Attention

Siyu Ding, Mingchuan Ma, Jiabo Tong, Xingrun Xing, Ziming Wang, Guoqi Li

Recent NVFP4 pretraining methods mainly target transformer linear layers, leaving optimizer states, optimizer arithmetic and attention underexplored in 4-bit pipelines. This critical gap blocks stable full-stack 4-bit pretraining, as the three core modules exhibit unique numerica…Read more