Cross-layer Attention Sharing for Pre-trained Large Language Models
This addresses efficiency bottlenecks for deploying large language models, though it is an incremental improvement over existing compression methods.
The paper tackles redundancy in attention mechanisms across layers of large language models by introducing LISA, which shares attention weights between layers using alignment networks and low-rank approximations, achieving 53%-84% reduction in redundant calculations while maintaining accuracy and improving throughput by up to 40.1%.
To enhance the efficiency of the attention mechanism within large language models (LLMs), previous works primarily compress the KV cache or group attention heads, while largely overlooking redundancy between layers. Our comprehensive analyses across various LLMs show that highly similar attention patterns persist within most layers. It's intuitive to reduce the redundancy by sharing attention weights across layers. However, further analysis reveals two challenges: (1) Directly sharing the weight matrix without carefully rearranging the attention heads proves to be ineffective; (2) Shallow layers are vulnerable to small deviations in attention weights. Driven by these insights, we introduce LISA, a lightweight substitute for self-attention in well-trained LLMs. LISA employs tiny feed-forward networks to align attention heads between adjacent layers and low-rank matrices to approximate differences in layer-wise attention weights. Evaluations encompassing 13 typical benchmarks demonstrate that LISA maintains high response quality in terms of accuracy and perplexity while reducing redundant attention calculations within 53%-84% of the total layers. Our implementations of LISA achieve a 6x compression of Q and K matrices within the attention mechanism, with maximum throughput improvements 19.5%, 32.3%, and 40.1% for LLaMA3-8B, LLaMA2-7B, and LLaMA2-13B, respectively.