Saved in:
Bibliographic Details
Main Authors: Kumar, Bhavesh, Jin, Roger, Quesnelle, Jeffrey
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2512.13728
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866911321831768064
author Kumar, Bhavesh
Jin, Roger
Quesnelle, Jeffrey
author_facet Kumar, Bhavesh
Jin, Roger
Quesnelle, Jeffrey
contents As language models scale to trillions of parameters, distributed training across many GPUs becomes essential, yet gradient synchronization over high-bandwidth, low-latency networks remains a critical bottleneck. While recent methods like Dion reduce per-step communication through low-rank updates, they synchronize at every step regardless of the optimization landscape. We observe that synchronization requirements vary dramatically throughout training: workers naturally compute similar gradients in flat regions, making frequent synchronization redundant, while high-curvature regions require coordination to prevent divergence. We introduce CurvaDion, which uses Relative Maximum Momentum Change (RMMC) to detect high-curvature regions requiring synchronization. RMMC leverages momentum dynamics which are already computed during optimization as a computationally tractable proxy for directional curvature, adding only $\mathcal{O}(d)$ operations per layer. We establish theoretical connections between RMMC and loss curvature and demonstrate that CurvaDion achieves 99\% communication reduction while matching baseline convergence across models from 160M to 1.3B parameters.
format Preprint
id arxiv_https___arxiv_org_abs_2512_13728
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle CurvaDion: Curvature-Adaptive Distributed Orthonormalization
Kumar, Bhavesh
Jin, Roger
Quesnelle, Jeffrey
Machine Learning
Artificial Intelligence
As language models scale to trillions of parameters, distributed training across many GPUs becomes essential, yet gradient synchronization over high-bandwidth, low-latency networks remains a critical bottleneck. While recent methods like Dion reduce per-step communication through low-rank updates, they synchronize at every step regardless of the optimization landscape. We observe that synchronization requirements vary dramatically throughout training: workers naturally compute similar gradients in flat regions, making frequent synchronization redundant, while high-curvature regions require coordination to prevent divergence. We introduce CurvaDion, which uses Relative Maximum Momentum Change (RMMC) to detect high-curvature regions requiring synchronization. RMMC leverages momentum dynamics which are already computed during optimization as a computationally tractable proxy for directional curvature, adding only $\mathcal{O}(d)$ operations per layer. We establish theoretical connections between RMMC and loss curvature and demonstrate that CurvaDion achieves 99\% communication reduction while matching baseline convergence across models from 160M to 1.3B parameters.
title CurvaDion: Curvature-Adaptive Distributed Orthonormalization
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2512.13728