DReSS: Data-driven Regularized Structured Streamlining for Large Language Models

TL;DR

DReSS introduces a data-driven regularization approach before pruning in large language models, effectively preserving important information and enabling more aggressive pruning with better performance and efficiency.

Contribution

The paper proposes a novel paradigm of regularize-then-prune for LLMs and introduces DReSS, a method that improves pruning effectiveness by transferring important information beforehand.

Findings

DReSS outperforms existing pruning methods at extreme pruning ratios.

DReSS reduces latency and increases throughput significantly.

Regularization before pruning preserves more model information.

Abstract

Large language models (LLMs) have achieved significant progress across various domains, but their increasing scale results in high computational and memory costs. Recent studies have revealed that LLMs exhibit sparsity, providing the potential to reduce model size through pruning techniques. However, existing pruning methods typically follow a prune-then-finetune paradigm. Since the pruned components still contain valuable information, their direct removal often leads to irreversible performance degradation, imposing a substantial computational burden to recover performance during finetuning. In this paper, we propose a novel paradigm that first applies regularization, then prunes, and finally finetunes. Based on this paradigm, we introduce DReSS, a simple and effective Data-driven Regularized Structured Streamlining method for LLMs. By leveraging a small amount of data to regularize the components to be pruned, DReSS explicitly transfers the important information to the remaining parts of the model in advance. Compared to direct pruning, this can reduce the information loss caused by parameter removal, thereby enhancing its language modeling capabilities. Experimental results demonstrate that DReSS significantly outperforms existing pruning methods even under extreme pruning ratios, significantly reducing latency and increasing throughput.