The Sharpness Disparity Principle in Transformers for Accelerating Language Model Pre-Training

Fuente: arXiv
Enregistré dans:
Détails bibliographiques
Auteurs principaux: Wang, Jinbo, Wang, Mingze, Zhou, Zhanpeng, Yan, Junchi, E, Weinan, Wu, Lei
Format: Preprint
Publié: 2025
Sujets:
Accès en ligne:
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866916793130418176
author Wang, Jinbo
Wang, Mingze
Zhou, Zhanpeng
Yan, Junchi
E, Weinan
Wu, Lei
author_facet Wang, Jinbo
Wang, Mingze
Zhou, Zhanpeng
Yan, Junchi
E, Weinan
Wu, Lei
contents Transformers consist of diverse building blocks, such as embedding layers, normalization layers, self-attention mechanisms, and point-wise feedforward networks. Thus, understanding the differences and interactions among these blocks is important. In this paper, we uncover a clear Sharpness Disparity across these blocks, which emerges early in training and intriguingly persists throughout the training process. Motivated by this finding, we propose Blockwise Learning Rate (LR), a strategy that tailors the LR to each block's sharpness, accelerating large language model (LLM) pre-training. By integrating Blockwise LR into AdamW, we consistently achieve lower terminal loss and nearly $2\times$ speedup compared to vanilla AdamW. We demonstrate this acceleration across GPT-2 and LLaMA, with model sizes ranging from 0.12B to 2B and datasets of OpenWebText, MiniPile, and C4. Finally, we incorporate Blockwise LR into Adam-mini (Zhang et al., 2024), a recently proposed memory-efficient variant of Adam, achieving a combined $2\times$ speedup and $2\times$ memory saving. These results underscore the potential of exploiting the sharpness disparity to improve LLM training.
format Preprint
id arxiv_https___arxiv_org_abs_2502_19002
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle The Sharpness Disparity Principle in Transformers for Accelerating Language Model Pre-Training
Wang, Jinbo
Wang, Mingze
Zhou, Zhanpeng
Yan, Junchi
E, Weinan
Wu, Lei
Machine Learning
Artificial Intelligence
Optimization and Control
Transformers consist of diverse building blocks, such as embedding layers, normalization layers, self-attention mechanisms, and point-wise feedforward networks. Thus, understanding the differences and interactions among these blocks is important. In this paper, we uncover a clear Sharpness Disparity across these blocks, which emerges early in training and intriguingly persists throughout the training process. Motivated by this finding, we propose Blockwise Learning Rate (LR), a strategy that tailors the LR to each block's sharpness, accelerating large language model (LLM) pre-training. By integrating Blockwise LR into AdamW, we consistently achieve lower terminal loss and nearly $2\times$ speedup compared to vanilla AdamW. We demonstrate this acceleration across GPT-2 and LLaMA, with model sizes ranging from 0.12B to 2B and datasets of OpenWebText, MiniPile, and C4. Finally, we incorporate Blockwise LR into Adam-mini (Zhang et al., 2024), a recently proposed memory-efficient variant of Adam, achieving a combined $2\times$ speedup and $2\times$ memory saving. These results underscore the potential of exploiting the sharpness disparity to improve LLM training.
title The Sharpness Disparity Principle in Transformers for Accelerating Language Model Pre-Training
topic Machine Learning
Artificial Intelligence
Optimization and Control
url https://arxiv.org/abs/2502.19002