基于动态Power迭代的大语言模型微调算法OA

With the widespread application of large language models(LLMs)across various domains,fine-tuning has become a critical method for adapting these models to specific tasks.Current mainstream fine-tuning approaches can be categorized into two types:partial fine-tuning and full fine-tuning.Although partial fine-tuning reduces computational overhead,it updates only a small subset of model parameters,which limits the performance of the fine-tuned model in complex task scenarios.In contrast,full fine-tuning overcomes this limitation by updating all model parameters,but it comes at the cost of significantly higher computational requirements and longer fine-tuning time.To address the challenges associated with full fine-tuning,this paper proposes a dynamic power iteration gradient low-rank projection algorithm(DPI-GLRP).Specifically,this method is based on the idea of rank-one matrix approximation,where the gradient matrix generated from the weight matrix of a large model during backpropagation is decomposed into multiple rank-one matrices.The top r eigenvectors are then computed using the Power iteration algorithm to construct the projection matrix.This approach not only overcomes the limitation of traditional Power iteration,which can extract only the dominant eigenvector,but also addresses the high time complexity and long fine-tuning duration associated with constructing projection matrices using singular value decomposition in previous studies.Furthermore,this paper analyzes the traditional Power iteration method and observes that it converges slowly when eigenvalues are closely distributed.To tackle this issue,this paper introduces a dynamic Power iteration algorithm that adaptively adjusts its iteration parameters to accelerate the computation of eigenvectors.This paper also provides theoretical proof that the proposed dynamic Power iteration achieves faster convergence compared with the traditional variant.Finally,experiments conducted on large models such as LLaMA and Qwen demonstrate that,compared with mainstream approaches like LoRA,the DPI-GLRP algorithm not only maintains or improves model performance,but also significantly reduces fine-tuning time,achieving up to an 80%reduction on average.