@nvidia's Jingfan Sun and Yang Xu share what it took to get SGLang running fast on Vera Rubin from day zero: finding bottlenecks and optimizing kernels and attention for large MoE models.
Practical lessons for anyone optimizing SGLang on new hardware.