Data-Free Pruning of Self-Attention Layers in LLMs
分析
This paper introduces Gate-Norm, a novel method for pruning self-attention layers in large language models (LLMs) without requiring any training data. The core idea revolves around the \
要点
引用
“Pruning $8$--$16$ attention sublayers yields up to $1.30\times$ higher inference throughput while keeping average zero-shot accuracy within $2\%$ of the unpruned baseline.”