Scale-Distribution Decoupling: Enabling Stable and Effective Training of Large Language Models

TL;DR

This paper introduces Scale-Distribution Decoupling (SDD), a novel method that stabilizes large language model training by decoupling weight scale and distribution, improving gradient stability and training efficiency.

Contribution

The paper proposes SDD, a new approach that explicitly separates scale and distribution in weight matrices to enhance training stability of large language models.

Findings

SDD stabilizes training across various LLM architectures.

SDD outperforms existing normalization techniques.

SDD is lightweight and compatible with current frameworks.

Abstract

Training stability is a persistent challenge in the pre-training of large language models (LLMs), particularly for architectures such as Post-Norm Transformers, which are prone to gradient explosion and dissipation. In this paper, we propose Scale-Distribution Decoupling (SDD), a novel approach that stabilizes training by explicitly decoupling the scale and distribution of the weight matrix in fully-connected layers. SDD applies a normalization mechanism to regulate activations and a learnable scaling vector to maintain well-conditioned gradients, effectively preventing $\textbf{gradient explosion and dissipation}$. This separation improves optimization efficiency, particularly in deep networks, by ensuring stable gradient propagation. Experimental results demonstrate that our method stabilizes training across various LLM architectures and outperforms existing techniques in different normalization configurations. Furthermore, the proposed method is lightweight and compatible with existing frameworks, making it a practical solution for stabilizing LLM training. Code is available at https://github.com/kaihemo/SDD.