arxivcs.PFcs.AIcs.AR2026-06-26
KernelSight-LM: A Kernel-Level LLM Inference Simulator
Xiteng Yao, Taeho Kim, Hengzhi Pei, Xinle Liu, Kyle Ulrich, Leonard Lausen, et al.
As large language models (LLMs) move into production serving, practitioners must rapidly evaluate inference performance across diverse hardware, models, and serving parameters to meet cost and latency targets. However, the end-to-end behavior of LLMs couples serving-layer policie…