Pretraining large language models with MXFP4

📰 ArXiv cs.AI

arXiv:2605.09825v1 Announce Type: cross Abstract: 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 of MXFP4 quantization in transformer training, progressively enabling FP4 across forward propagation (Fprop), activation gradients (Dgrad), and weight gradients (Wgrad) while holding all other factors fixed. In full pretraining of Llama 3.1-8B on the C

Published 12 May 2026

Full Article

Title: Pretraining large language models with MXFP4

Abstract:
arXiv:2605.09825v1 Announce Type: cross Abstract: 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 of MXFP4 quantization in transformer training, progressively enabling FP4 across forward propagation (Fprop), activation gradients (Dgrad), and weight gradients (Wgrad) while holding all other factors fixed. In full pretraining of Llama 3.1-8B on the C
Read full paper → ← Back to Reads