About performance when serving CoT/reasoning models

[Hygge02]

Hi! Thanks for your great work!

Have you tested the performance when CoT/reasoning models such as DeepSeek-Distill-Series are served by LServe? In the LServe paper, it's mentioned that LServe aims at accelerating both prefilling and decoding stage. But it seems that there are no results comming from reasoning benchmark such as AIME 2024, MATH-500 etc.

Thanks for your reply!