Checklist

• 1. If the issue you raised is not a feature but a question, please raise a discussion at https://github.com/sgl-project/sglang/discussions/new/choose Otherwise, it will be closed.
• 2. Please use English, otherwise it will be closed.

Motivation

Currently all the bf16 gemm kernels in SGLang are called from torch.linear, with underlying Cublas kernels. On B200 CuteDSL might provide bf16 gemm kernels faster than Cublas.

https://github.com/Dao-AILab/quack/blob/main/quack/dense_gemm_sm100.py

Related resources

No response