LLM-BIP: Structured Pruning for Large Language Models with Block-Wise Forward Importance Propagation

TL;DR

This paper introduces LLM-BIP, a block-wise importance propagation method for structural pruning of large language models, improving accuracy and reducing perplexity by accurately evaluating connection importance during pruning.

Contribution

The paper proposes a novel block-wise importance score propagation technique for structural pruning, addressing inaccuracies in existing global and layer-wise methods.

Findings

Achieves 3.26% higher accuracy on reasoning tasks

Reduces perplexity by 14.09 on WikiText2

Reduces perplexity by 68.76 on PTB

Abstract

Large language models (LLMs) have demonstrated remarkable performance across various language tasks, but their widespread deployment is impeded by their large size and high computational costs. Structural pruning is a prevailing technique used to introduce sparsity into pre-trained models and facilitate direct hardware acceleration during inference by removing redundant connections (structurally-grouped parameters), such as channels and attention heads. Existing structural pruning approaches often employ either global or layer-wise pruning criteria; however, they are hindered by ineffectiveness stemming from inaccurate evaluation of connection importance. Global pruning methods typically assess component importance using near-zero and unreliable gradients, while layer-wise pruning approaches encounter significant pruning error accumulation issues. To this end, we propose a more accurate pruning metric based on the block-wise importance score propagation, termed LLM-BIP. Specifically, LLM-BIP precisely evaluates connection importance by gauging its influence on the respective transformer block output, which can be efficiently approximated in a single forward pass through an upper bound derived from the assumption of Lipschitz continuity. We evaluate the proposed method using LLaMA-7B, Vicuna-7B, and LLaMA-13B across common zero-shot tasks. The results demonstrate that our approach achieves an average of 3.26% increase in accuracy for common reasoning tasks compared to previous best baselines. It also reduces perplexity by 14.09 and 68.76 on average for the WikiText2 dataset and PTB dataset, respectively.