Pruning via Merging: Compressing LLMs via Manifold Alignment Based Layer Merging

TL;DR

This paper introduces MKA, a novel layer merging technique using manifold learning and information bottleneck to compress large language models efficiently while maintaining performance.

Contribution

The paper presents MKA, a new model compression method that merges similar layers based on manifold alignment, outperforming traditional pruning and working well with quantization.

Findings

MKA achieves a 43.75% compression ratio on Llama3-8B with only 2.82% performance loss.

MKA outperforms traditional pruning methods in preserving model accuracy.

Coupling MKA with quantization yields even higher compression ratios.

Abstract

While large language models (LLMs) excel in many domains, their complexity and scale challenge deployment in resource-limited environments. Current compression techniques, such as parameter pruning, often fail to effectively utilize the knowledge from pruned parameters. To address these challenges, we propose Manifold-Based Knowledge Alignment and Layer Merging Compression (MKA), a novel approach that uses manifold learning and the Normalized Pairwise Information Bottleneck (NPIB) measure to merge similar layers, reducing model size while preserving essential performance. We evaluate MKA on multiple benchmark datasets and various LLMs. Our findings show that MKA not only preserves model performance but also achieves substantial compression ratios, outperforming traditional pruning methods. Moreover, when coupled with quantization, MKA delivers even greater compression. Specifically, on the MMLU dataset using the Llama3-8B model, MKA achieves a compression ratio of 43.75% with a minimal performance decrease of only 2.82\%. The proposed MKA method offers a resource-efficient and performance-preserving model compression technique for LLMs.