Wanda++ introduces a novel pruning framework that uses regional gradients to identify and remove unimportant weights in Large Language Models without significant performance loss.

• Achieves superior pruning results without requiring full-model sparsity-aware fine-tuning
• Leverages decoder-block-level regional gradients to improve pruning score accuracy
• Delivers better model performance at the same sparsity level compared to state-of-the-art methods
• Enables faster inference with minimal impact on model capabilities

This engineering breakthrough addresses a critical challenge in LLM deployment: balancing computational efficiency with model performance, making advanced AI more accessible and cost-effective for real-world applications.

Original Paper: Wanda++: Pruning Large Language Models via Regional Gradients