Researchers developed a novel framework that strategically reduces LLM size while preserving performance through a structured two-stage pruning approach.

• Width Pruning: First stage removes entire neurons while preserving connectivity in Feed-Forward Networks
• Depth Pruning: Second stage identifies and removes less important Transformer blocks
• Balanced Approach: Combines techniques to maintain model integrity while significantly reducing parameters
• Engineering Impact: Enables more efficient deployment of LLMs in resource-constrained environments

This research advances model optimization techniques essential for practical LLM deployment in production systems where computational resources are limited.

2SSP: A Two-Stage Framework for Structured Pruning of LLMs