A Large Model Pruning and Heterogeneous Replacement Scheme Based on Consecutive Layer Influence

• 1、Computer Science School，Beijing University of Posts and Telecommunications，Beijing 100876

Abstract：This paper aims to propose an efficient large model compression scheme that significantly reduces the model parameter size and improves inference speed, while maintaining the original accuracy or with only slight degradation. To achieve this, the paper introduces a pruning and heterogeneous replacement approach based on the influence of consecutive layers in the large model. The method first calculates the influence of all consecutive layers in the model and then selects one or more layers with the least influence for heterogeneous replacement based on related algorithms. Finally, the compressed model is fine-tuned through knowledge distillation to recover its performance, resulting in a lightweight model. This study also conducts relevant experiments to evaluate the lightweight model, and the final results show that its performance is close to that of the original large model. The research demonstrates the effectiveness of the proposed pruning and heterogeneous replacement method based on consecutive layer influence, achieving a balance between model acceleration and accuracy retention, significantly reducing inference latency while maintaining model precision.

Keywords： Computer Science and Technology Large Model Compression Model Pruning Knowledge Distillation