CommBench: Can LLMs Write Correct and Efficient GPU Communication Code?

Shuang Ma, Yuyi Li, Yihan Zhang +12

Training and serving large language models (LLMs) rely heavily on high-performance GPU communication, yet implementing efficient GPU communication primitives requires deep expertis…