Qinghai Guo

Adaptive Depth Sparse Framework: Similarity-Driven Resource Allocation for Pre-Trained LLMs

Yidu Wu, Xiang Wang, Kejie Zhao, Zhangchi Wang, Qinghai Guo, Xiaoying Tang

Large language models (LLMs) achieve strong generation and reasoning performance, but the Transformer architecture incurs high inference cost. Existing acceleration methods often rely on task-specific fine-tuning or training from scratch, increasing adaptation cost and limiting c…Read more