Dynamic Activation Pitfalls in LLaMA Models: An Empirical Study
Chi Ma, Mincong Huang, Chao Wang, Yujie Wang, Lei Yu
TL;DR
This paper empirically analyzes dynamic activation in LLaMA models, focusing on non-ReLU activations and high sparsity scenarios. It combines threshold-based sparsification in MLPs (CETT target $0.2$), dynamic head masking in attention, KV-cache skipping, and sparsity predictors to assess speedups and accuracy. The findings reveal systematic underperformance of current dynamic activation schemes relative to ReLU baselines, driven by activation-prediction complexity, insufficient sparsity for non-ReLU activations, and information loss from KV-cache skipping. The work provides roadmaps for layer-specific thresholds, more capable predictors, and KV-cache-aware strategies to guide future sparsity design for large LLMs.
Abstract
In this work, we systematically investigate the efficacy of dynamic activation mechanisms within the LLaMA family of language models. Despite the potential of dynamic activation methods to reduce computation and increase speed in models using the ReLU activation function, our empirical findings have uncovered several inherent pitfalls in the current dynamic activation schemes. Through extensive experiments across various dynamic activation strategies, we demonstrate that LLaMA models usually underperform when compared to their ReLU counterparts, particularly in scenarios demanding high sparsity ratio. We attribute these deficiencies to a combination of factors: 1) the inherent complexity of dynamically predicting activation heads and neurons; 2) the inadequate sparsity resulting from activation functions; 3) the insufficient preservation of information resulting from KV cache skipping. Our analysis not only sheds light on the limitations of dynamic activation in the context of large-scale LLaMA models but also proposes roadmaps for enhancing the design of future sparsity schemes.