Pretraining large language models with MXFP4 on Native FP4 Hardware

Why does full-pipeline FP4 training of large language models often diverge, even when forward activations and activation gradients remain stable? We address this question through a controlled study…