Dispatch Implementation for Expert Parallelism

[Blueblack319]

I’m currently using SGLang v0.4.10 to run large MoE models such as Qwen3MoE and GLM-4.5 on a multi-GPU system.
After analyzing the executed kernels, I found that the first All2All (Dispatch) step is implemented as an AllGather operation.
Could you explain why SGLang implements Dispatch as AllGather instead of using All2All (i.e., all_to_all_single)?

[MaoZiming]

• 1. I hope to know how the dispatch and combine phase under --moe-a2a-backend none is implemented.