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 dsv3_router_gemm outputs kernel with fp32 precision. When applying it to deepseek-r1 fp4, the output is converted to bf16 to avoid bug. However, these extra conversions will cause performance drop when bs >= 4.

We should modify router_gemm kernel and make it output bf16 tensor directly. Then dsv3_router_gemm kernel can be applied to bs >=4

Ref: #7627, #7677

Related resources

No response