Unusually high Communication cost on DeepSeek pure EP

[jorjiang]

according to the simulation: for deepseek v3 671b, dp=ep=128
the moe dispatch and combine took up over 90% latency, which is unreasonable

[tianhaox]

This is because for trtllm non-wideep, it's calling NCCL for the all2all communication. TRTLLM based wideep path is not supported yet. I will suggest trying sglang h100/h200 wideep (ep<=64) for ep>num_gpu_per_node,

[tianhaox]

sglang's pre/post comm is alltogether in pre_dispatch.