Your LR Schedule May Not Be Principled: Effective Learning Rate Scheduling in Weight-Normalized Training
Qifan Yu, Xinyu Ma, Xingyan Bin, Di He
原文摘要
Weight-normalized (WN) training is emerging as a compelling alternative to conventional weight decay (WD), with recent studies reporting substantial pretraining loss gains. We find that the effective learning rate (ELR) trajectory accounts for most of the loss dynamics difference between WN and WD. Matching blockwise ELR trajectories step by step closely aligns their loss curves. Schedule design should therefore target ELR rather than the nominal learning rate (LR). Building on this principle, we model directional-alignment dynamics under WN and derive $1/\sqrt{t}$ decay as the optimal schedule. For finite-horizon pretraining, we propose AIR (Affine Inverse-square-Root decay) to match prescribed peak and terminal LRs. Experiments on 128B- and 500B-token Mixture-of-Experts pretraining under WN show that AIR outperforms cosine and alternative schedules in loss, downstream performance, and hyperparameter robustness, reducing terminal training loss by 0.01 relative to cosine. We further identify late-stage scheduling as the primary source of this gain through controlled stage interventions, and demonstrate that a shared late schedule progressively attenuates alignment differences left by earlier LRs. Finally, transferring AIR's ELR trajectory to WD reproduces the same improved loss dynamics, demonstrating that these gains come from ELR scheduling itself and extend beyond fixed-norm training.
cs.LG