Although Large Language Models (LLMs) based on the Transformer architecture have developed rapidly in Natural Language Processing (NLP), their massive parameter counts and high computational costs severely constrain practical deployment.Existing model compression techniques face two critical challenges, namely the reliance on gradient information for weight importance assessment and the necessity of full parameter updating after pruning, which increases training costs.To address these challenges, this paper proposes a structured pruning method guided by Low-Rank Adaptation (LoRA) for importance assessment, aiming to achieve efficient model compression and fine-tuning.During pruning, the proposed method innovatively utilizes gradient information of the low-rank matrices in the LoRA module to approximate the importance of original parameters.By avoiding the direct gradient calculation of original weights, this method effectively enhances computational efficiency and hardware compatibility.In the performance recovery phase, lightweight parameter updates are performed via the LoRA module, thereby reducing the resource consumption associated with full-parameter fine-tuning.Furthermore, this paper formulates pruning as a multi-objective optimization problem and introduces an exponential sparsity regulation strategy to achieve a dynamic balance among performance, resource consumption, and compression rate.Evaluations on the General Language Understanding Evaluation (GLUE) benchmark demonstrate that, without pre-fine-tuning, the proposed method maintains model performance while reducing GPU memory usage, showing superiority over existing pruning strategies.Specifically, the proposed method reduces GPU memory usage by up to 68.9% compared to traditional approaches.
本文提出了一种基于LoRA评估重要性的结构化剪枝方法,旨在在不引入大量计算与显存开销的前提下,实现高效的模型参数压缩与性能恢复。具体地,如图2所示,本文通过引入轻量的低秩可训练矩阵 A 和 B,在微调过程中利用其梯度信息评估原始权重的重要性,来替代传统剪枝方法中对权重梯度的依赖。其次,引入约束函数并采用指数稀疏率调节策略逐步剪枝MHA和FFN模块的冗余参数,每轮剪枝后通过对LoRA参数的更新即可快速恢复模型性能,无需全量参数更新,显著提升训练效率与部署适应性。
为了避免一次性剪枝对模型带来不可恢复的性能损失和提升剪枝策略的灵活性与收敛性能,本文引入一种基于指数增长的稀疏率调节函数,用于动态控制剪枝强度的变化过程。该策略能使模型能够在训练过程中逐渐适应参数的减少,提高模型的稳定性和鲁棒性以及通过在训练过程中逐渐增大压缩力度,使得模型在可恢复范围内逐渐满足目标函数。同时,对LoRA的参数 A 和 B 进行更新以恢复模型性能。具体形式如下所示:
VaswaniA, ShazeerN, ParmarN, et al.Attention is all you need[J].Adv Neural Inf Process Syst, 2017,30:5998-6008.
[2]
DevlinJ, ChangM W, LeeK, et al.BERT: Pre-training of deep bidirectional transformers for language understanding[C]//Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 2019: 4171-4186.
[3]
LiuY, OttM, GoyalN, et al.RoBERTa: A robustly optimizedBERT pretraining approach [PP/OL].[2025-06-28].
[4]
BrownT, MannB, RyderN, et al.Language models are few-shot learners[J].Adv Neural Inf Process Syst, 2020, 33: 1877-1901.
[5]
WangA, SinghA, MichaelJ, et al.GLUE: A multi-task benchmark and analysis platform for natural language understanding[C]//Proceedings of the 2018 EMNLP Workshop Blackbox NLP: Analyzing and Interpreting Neural Networks for NLP, 2018: 353-355.
[6]
DosovitskiyA, BeyerL, KolesnikovA, et al.An image is worth 16×16 words: Transformers for image recognition at scale[PP/OL].[2025-06-28].
[7]
ShihK-H, ChiuC-T, LinJ-A, et al.Real-time object detectionwith reduced region proposal network via multi-feature concatenation[J].IEEE Trans Neural Netw Learn Syst, 2020, 31(6): 2164-2173.
[8]
TangY, WangY, GuoJ, et al.A survey on transformer compression[PP/OL].[2025-06-28].
[9]
LinJ, TangJ, TangH, et al.AWQ: Activation-aware weight quantization for on-device LLM compression and acceleration[J].GetMobile: Mobile Comp and Comm, 2025, 28(4): 12-17.
[10]
AnagnostidisS, BiggioL, HofmannT, et al.Dynamic context pruning for efficient and interpretable autoregressive transformers[J].Adv Neural Inf Process Syst, 2023, 36: 65202-65223.
[11]
VoitaE, TalbotD, MoiseevF, et al.Analyzing multi-head self-attention: Specialized heads do the heavy lifting, the rest can bepruned[C]//Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, 2019: 5797-5808.
[12]
WangZ, HongW, TanY-P, et al.Pruning 3D filters for accelerating 3D ConvNets[J].IEEE Trans Multimed, 2020, 22(8):2126-2137.
[13]
WangH, ZhangW Q.Unstructured pruning and low rank factorisation of self-supervised pre-trained speech models[J].IEEE J Sel Top Signal Process, 2024, 18(6): 1046-1058.
[14]
HanS, MaoH, DallyW J.Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding[PP/OL].[2025-06-28].
[15]
MichelP, LevyO, NeubigG.Are sixteen heads really better than one?[C]//Proceedings of the 33rd International Conference on Neural Information Processing Systems, 2019: 14037-14047.
[16]
KangH J.Accelerator-aware pruning for convolutional neural networks[J].IEEE Trans Circuits Syst Video Technol, 2020, 30(7):2093-2103.
HuE J, ShenY, WallisP, et al.Lora: Low-rank adaptation of large language models [PP/OL].[2025-06-28].
[19]
FangG, MaX, WangX.LLM-pruner: On the structural pruningof large language models[C]//Advances in Neural Information Processing Systems 36, 2023: 21702-21720.
[20]
LeCunY, DenkerJ, SollaS.Optimal brain damage [C]//Proceedings of the 3rd International Conference on Neural Information Processing Systems, 1989: 598-605.
[21]
LeeJ, ParkS, MoS, et al.Layer-adaptive sparsity for the magnitude-based pruning[PP/OL].[2025-06-28].
[22]
MolchanovP, MallyaA, TyreeS, et al.Importance estimation for neural network pruning[C]//2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019: 11256-11264.
[23]
KwonW, KimS, MahoneyM W, et al.A fast post-training pruning framework for transformers[J].Adv Neural Inf Process Syst, 2022, 35: 24101-24116.
[24]
LiuZ, LiF, LiG, et al.EBERT: Efficient BERT inference with dynamic structured pruning[C]//Findings of the Association forComputational Linguistics: ACL-IJCNLP 2021, 2021: 4814-4823.
[25]
XiaM, ZhongZ, ChenD.Structured pruning learns compact and accurate models[C]//Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics, 2022: 1513-1528.
[26]
WangZ, WohlwendJ, LeiT.Structured pruning of large language models[C]//Proceedings of the 2020 Conference on EmpiricalMethods in Natural Language Processing (EMNLP), 2020: 6151-6162.
[27]
LagunasF, CharlaixE, SanhV, et al.Block pruning for faster transformers[C]//Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, 2021: 10619-10629.
[28]
IlhanF, SuG, TekinS F, et al.Resource- efficient transformerpruning for finetuning of large models [C]//2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024: 16206-16215.
[29]
SajjadH, DalviF, DurraniN, et al.Poor man’s bert: Smaller and faster transformer models[PP/OL].[2025-06-28].
[30]
LinZ, LiuJ, YangZ, et al.Pruning redundant mappings in transformer models via spectral-normalized identity prior[C]//Findings of the Association for Computational Linguistics: EMNLP 2020, 2020: 719-730.
[31]
LimH, LeeJ, ChoiD W.PGB: One-shot pruning for BERT via weight grouping and permutation[J].Jair, 2026, 85: 1-20.