NVIDIA engineers describe tuning SGLang for Vera Rubin and large MoE models

NVIDIA engineers explain how they brought SGLang to Vera Rubin, according to SGLang's account on X. Lee Nau and Akhil Goel describe finding performance bottlenecks and tuning kernels and attention for large mixture-of-experts (MoE) models.

A separate post from SGLang's account has NVIDIA's Jingfan Sun and Yang Xu describing the same work, including getting SGLang running on Vera Rubin from day zero.

Both posts are promotional summaries published by the SGLang project, and they give no benchmark figures or performance results.