1
Tile-Level Activation Overlap for Efficient LLM Inference
突破LLM推理瓶颈：块级激活重叠技术显著降低显存占用，提升推理吞吐与效率。
arXiv:2607.02521v1 Announce Type: cross Abstract: SwiGLU is the dominant MLP activation in modern large language models, yet its intermediate tensor m…