WiSparse: Boosting LLM Inference Efficiency with Weight-Aware Mixed Activation Sparsity

TL;DR

WiSparse introduces a weight-aware, mixed-granularity activation sparsity method that significantly improves inference efficiency in large language models without retraining, maintaining high accuracy at high sparsity levels.

Contribution

This paper presents WiSparse, a novel training-free sparsity approach that combines activation and weight information with adaptive allocation to enhance LLM inference efficiency.

Findings

At 50% sparsity, WiSparse retains 97% of Llama3.1's dense performance.

WiSparse achieves a 21.4% inference speedup over baseline methods.

It surpasses existing activation sparsity techniques in accuracy preservation.

Abstract

Large Language Models (LLMs) offer strong capabilities but incur high inference costs due to dense computation and memory access. Training-free activation sparsity is a promising approach for efficient LLM inference, yet existing methods often rely solely on activation information and uniform sparsity ratios. This overlooks the critical interplay with weights and inter-block sensitivity variation, leading to suboptimal performance. We identify two key phenomena in modern LLMs: 1) less significant activations may align with highly important weights, and 2) sparsity sensitivity varies non-monotonically across model blocks. We propose Weight-aware Mixed-Granularity Training-free Activation Sparsity (WiSparse), which leverages both activation and weight information for adaptive sparsity allocation. Specifically, we introduce a weight-aware mechanism integrating activation magnitudes with precomputed weight norms to accurately identify salient channels. This is combined with a mixed-granularity allocation scheme: a global budget is distributed across blocks via evolutionary search to protect sensitive regions, then refined within blocks to minimize reconstruction error. We improve sparse kernels and demonstrate effectiveness on three representative models. Notably, at 50% sparsity, WiSparse preserves 97% of Llama3.1's dense performance, surpassing the strongest baseline by 2.23 percentage points while achieving a 21.4% acceleration in end-to-end inference speed. Our research advances the limits of training-free approaches for efficient LLM inference, pushing the boundaries of achievable speedup without training.