Understand and Accelerate Memory Processing Pipeline for Large Language Model Inference
理解并加速大型语言模型推理的内存处理流水线
AI总结 本文通过将稀疏注意力、检索增强生成和压缩上下文内存等优化统一为四步内存处理流水线,识别出22%-97%的内存处理开销,并提出使用GPU-FPGA异构系统加速该流水线,实现最高2.2倍加速和4.7倍能效提升。
Comments Accepted by ICML 2026. Code: https://github.com/OswaldHe/HeteroLLM