arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2604.01563cs.AIcs.LG
收藏

你的优化器是否在意你如何归一化?LLM训练中的归一化-优化器耦合

Does Your Optimizer Care How You Normalize? Normalization-Optimizer Coupling in LLM Training

Georgia Institute of Technology(佐治亚理工学院)

Abdelrahman Abouzeid

AI总结:

研究发现归一化层与优化器在LLM训练中存在交互影响,动态Erf在Muon优化器下表现下降,通过调整参数可恢复性能,揭示归一化方法对优化器的敏感性。

摘要:

在LLM训练中,归一化层和优化器通常被视为独立设计选择。在10亿参数和1000次训练步骤的3x2因子实验中,我们发现这一假设可能不成立:动态Erf(Derf)与Muon优化器存在显著负交互,其与RMSNorm的差距从AdamW下的+0.31纳特增加到Muon下的+0.97纳特,大约大三倍。动态Tanh(DyT)作为有界归一化器控制,未表现出此类惩罚。我们的证据表明,Erf在Muon更快的谱范数增长下存在两种失效模式:饱和(损失性压缩)和尺度盲(丢弃激活幅度)。一个EMA混合方法重新引入运行尺度估计可恢复约84%的差距。此外,减少Derf的alpha值(从其发布的默认值0.5到0.3)可恢复约80%,通过保持Erf在近线性区域,其中它近似保留相对尺度;此设置并非Chen & Liu(2025)发布的默认值。使用Derf发布的默认alpha值与Muon会产生0.66纳特的交互惩罚,但不会产生NaN或发散,使失败在短试点运行中容易被遗漏。

英文摘要:

In LLM training, normalization layers and optimizers are typically treated as independent design choices. In a 3x2 factorial at 1B parameters and 1000 training steps, we show this assumption can fail: Dynamic Erf (Derf; Chen & Liu, 2025) suffers a large negative interaction with Muon (Jordan, 2024), with its gap to RMSNorm growing from +0.31 nats under AdamW to +0.97 under Muon, approximately three times larger. Dynamic Tanh (DyT; Zhu et al., 2025), included as a bounded-normalizer control, shows no such penalty. Our evidence points to two failure modes of erf under Muon's faster spectral-norm growth: saturation (lossy compression) and scale blindness (discarding activation magnitude). An EMA-blend that reintroduces running scale estimates recovers ~84% of the gap. Separately, reducing Derf's alpha from its published default (0.5 to 0.3) recovers ~80% by keeping erf in its near-linear regime, where it approximately preserves relative scale; this setting is not the published default of Chen & Liu (2025). Using Derf's published default alpha with Muon incurs a 0.66-nat interaction penalty without producing NaNs or divergence, making the failure easy to miss in short pilot runs.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑