Support qwen3 deepep

[sleepcoo]

Motivation

Support qwen3's deepep. For now, we've simply copied the deepep code from DS, and the accuracy test has passed.

TODO

• test bf16 compatibility

Test Command

 python3 -m sglang.launch_server --model-path /workdir/huggingface.co/Qwen/Qwen3-235B-A22B-FP8/ --tp 4 --trust-remote --enable-torch-compile --torch-compile-max-bs 8 --host 0.0.0.0 --port 8418 --reasoning-parser qwen3 --tool-call-parser qwen25 --enable-deepep-moe --deepep-mode normal 

gsm8k Accuracy: 0.970

[yizhang2077]

Do we need raise error for bf16 when enable deepep?

[sleepcoo]

Do we need raise error for bf16 when enable deepep?

I'm not sure, it's necessary? @zhyncs @ch-wan

[ch-wan]

Do we need raise error for bf16 when enable deepep?

I'm not sure, it's necessary? @zhyncs @ch-wan

@sleepcoo How about raising NotImplementedError if it is not urgent?

some bus

[MtFitzRoy]

When can we support qwen3 bf16 with deepep?

[yizhang2077]

When can we support qwen3 bf16 with deepep?

I think deepep currently do not support bf16 dispatch?

[MtFitzRoy]

@yizhang2077 Seems DeepEP already support bf16 dispatch https://github.com/deepseek-ai/DeepEP?tab=readme-ov-file#roadmap

[yizhang2077]

@yizhang2077 Seems DeepEP already support bf16 dispatch https://github.com/deepseek-ai/DeepEP?tab=readme-ov-file#roadmap

I think normal mode is unmentioned

[MtFitzRoy]

@yizhang2077 For normal mode dispatch, I think sglang code already support bf16.