[Feature] support ep for DeepSeek V3

[zhyncs]

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

The code for EP and block wise FP8 required by V3 is available separately. The task is to integrate block wise FP8 into the current DeepSeek V2 EP, based on the previous integration of Fused MoE with block wise FP8.

ref

https://github.com/sgl-project/sglang/tree/main/python/sglang/srt/layers/moe/ep_moe

#2575

Related resources

No response

[annp0]

Hi @zhyncs , I'd be more than happy to take a look! :)

[zhyncs]

Hi @zhyncs , I'd be more than happy to take a look! :)

Thanks!!

[lixiaolx]

@zhyncs, When can this task be completed? Is there an approximate time?

[xinji1]

Hi @zhyncs we're working on it too. Will release the available codes asap and create a PR.

[zhyncs]

Hi @xinji1 Thanks! Please join the slack channel https://slack.sglang.ai

[zhyncs]

FYI The president of Meituan @sleepcoo will take over this feature. Cheers!

[YEXINGZHE54]

expect this feature

[sleepcoo]

expect this feature

On the way on the way