Dissecting Outlier Dynamics in LLM NVFP4 Pretraining

TL;DR

This paper analyzes outlier behaviors during NVFP4 pretraining of large language models, identifying architectural sources of outliers and proposing a novel online compensation method to improve training accuracy.

Contribution

It provides a detailed longitudinal analysis of outlier dynamics in NVFP4 training and introduces HCP and CHON to mitigate outlier effects and close the loss gap.

Findings

Outliers evolve from transient spikes to persistent hot channels during training.

Linear Attention reduces heavy tails but still exhibits block-level spikes.

CHON training recipe significantly reduces the loss gap to BF16.

Abstract

Training large language models using 4-bit arithmetic enhances throughput and memory efficiency. Yet, the limited dynamic range of FP4 increases sensitivity to outliers. While NVFP4 mitigates quantization error via hierarchical microscaling, a persistent loss gap remains compared to BF16. This study conducts a longitudinal analysis of outlier dynamics across architecture during NVFP4 pretraining, focusing on where they localize, why they occur, and how they evolve temporally. We find that, compared with Softmax Attention (SA), Linear Attention (LA) reduces per-tensor heavy tails but still exhibits persistent block-level spikes under block quantization. Our analysis attributes outliers to specific architectural components: Softmax in SA, gating in LA, and SwiGLU in FFN, with "post-QK" operations exhibiting higher sensitivity to quantization. Notably, outliers evolve from transient spikes early in training to a small set of persistent hot channels (i.e., channels with persistently large magnitudes) in later stages. Based on these findings, we introduce Hot-Channel Patch (HCP), an online compensation mechanism that identifies hot channels and reinjects residuals using hardware-efficient kernels. We then develop CHON, an NVFP4 training recipe integrating HCP with post-QK operation protection. On GLA-1.3B model trained for 60B tokens, CHON reduces the loss gap to BF16 from 0.94% to 0.58% while maintaining downstream accuracy.