Saved in:
Bibliographic Details
Main Authors: Kong, Boao, Jia, Weichen, Zhang, Engao, Li, Guohong, Dong, Yonghan, Wang, Yao, Wang, Yaoyuan, Peng, Yunke, Yuan, Kun
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2606.00539
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866917550205435904
author Kong, Boao
Jia, Weichen
Zhang, Engao
Li, Guohong
Dong, Yonghan
Wang, Yao
Wang, Yaoyuan
Peng, Yunke
Yuan, Kun
author_facet Kong, Boao
Jia, Weichen
Zhang, Engao
Li, Guohong
Dong, Yonghan
Wang, Yao
Wang, Yaoyuan
Peng, Yunke
Yuan, Kun
contents Training stability is a key bottleneck in low-precision language model training: efficient low-cost paths can still produce short-lived numerical risks at a small set of operators. We formulate this as runtime stability control and present Gradient Norm-to-Mean Ratio (GNMR), a lightweight controller that compares each recoverable unit's current gradient norm with its historical mean. Together with $Δ$-GNMR for abrupt short-window increases, GNMR maps local risk signals to bounded recovery actions under a hard $\mathrm{maxO}$ budget and a short lock interval, without changing the numerical format, kernel, or backend recipe. Across activation-quantization stress, DeepSeek-style recipe-level training, and LLaMA-2 13B fine-tuning, GNMR preserves high-fidelity quality with sparse, budgeted recovery. These results support GNMR as a backend-agnostic controller to improve low-precision training stability while preserving low-cost execution.
format Preprint
id arxiv_https___arxiv_org_abs_2606_00539
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle GNMR: Runtime Stability Control for Low-Precision Large Language Model Training
Kong, Boao
Jia, Weichen
Zhang, Engao
Li, Guohong
Dong, Yonghan
Wang, Yao
Wang, Yaoyuan
Peng, Yunke
Yuan, Kun
Machine Learning
Optimization and Control
Training stability is a key bottleneck in low-precision language model training: efficient low-cost paths can still produce short-lived numerical risks at a small set of operators. We formulate this as runtime stability control and present Gradient Norm-to-Mean Ratio (GNMR), a lightweight controller that compares each recoverable unit's current gradient norm with its historical mean. Together with $Δ$-GNMR for abrupt short-window increases, GNMR maps local risk signals to bounded recovery actions under a hard $\mathrm{maxO}$ budget and a short lock interval, without changing the numerical format, kernel, or backend recipe. Across activation-quantization stress, DeepSeek-style recipe-level training, and LLaMA-2 13B fine-tuning, GNMR preserves high-fidelity quality with sparse, budgeted recovery. These results support GNMR as a backend-agnostic controller to improve low-precision training stability while preserving low-cost execution.
title GNMR: Runtime Stability Control for Low-Precision Large Language Model Training
topic Machine Learning
Optimization and Control
url https://arxiv.org/abs/2606.00539