Large Language Models (LLMs) pruning seeks to remove unimportant weights for inference speedup with minimal performance impact. However, existing methods often suffer from performance loss without full-model sparsity-aware fine-tuning. This paper presents Wanda++, a novel pruning framework that outperforms the state-of-the-art methods by utilizing decoder-block-level \textbf{regional} gradients. Specifically, Wanda++ improves the pruning score with regional gradients for the first time and proposes an efficient regional optimization method to minimize pruning-induced output discrepancies between the dense and sparse decoder output. Notably, Wanda++ improves perplexity by up to 32\% over Wanda in the language modeling task and generalizes effectively to downstream tasks. Further experiments indicate our proposed method is orthogonal to sparsity-aware fine-tuning, where Wanda++ can be combined with LoRA fine-tuning to achieve a similar perplexity improvement as the Wanda method. The proposed method is lightweight, pruning a 7B LLaMA model in under 10 minutes on a single NVIDIA H100 GPU.
View on arXiv@article{yang2025_2503.04992,
title={ Wanda++: Pruning Large Language Models via Regional Gradients },
author={ Yifan Yang and Kai Zhen and Bhavana Ganesh and Aram Galstyan and Goeric Huybrechts and Markus Müller and Jonas M. Kübler and Rupak Vignesh Swaminathan and Athanasios Mouchtaris and Sravan Babu Bodapati and Nathan Susanj and Zheng Zhang and Jack FitzGerald and Abhishek Kumar },
journal={arXiv preprint arXiv:2503.04992},
year={ 2025 }
}