[vLLM-ATOM] Enable DBO for vLLM plugin

[kliuae]

Motivation

This PR enables DBO for vLLM-ATOM. This PR requires DP+EP enablement and currently contains changes from the enablement PR.

Technical Details

Test Plan

Test Result

deepseek-ai/DeepSeek-R1-0528 DP8+EP+DBO

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
gsm8k	3	flexible-extract	5	exact_match	_	0.953	_	0.0058
		strict-match	5	exact_match	_	0.950	_	0.0060

openai/gpt-oss-120b DP2+EP+DBO

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
gsm8k	3	flexible-extract	5	exact_match	_	0.4723	_	0.0138
		strict-match	5	exact_match	_	0.3237	_	0.0129

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[zejunchen-zejun]

this hard assert still exist in vLLM 0.22.0 version?

[kliuae]

Yes in vLLM v0.22.0 this still exists

[zejunchen-zejun]

We restore the all2all backend to avoid vllm create another mori all2all manager, after that, the sys will have 2 mori all2all managers, while it could not happen as there could be an unsupported error when specifying --all2all-backend=mori --enable-dbo because it is not supported in vllm for DBO for now.

Could you have a recipe about atom-vllm DBO usage? The users may specify a mori all2all manager when launch vllm server, but got error, what argument should atom-vllm frontend user specify

[kliuae]

Actually the restore_backend here is set to falling back to AgRs on vLLM side for all cases. So when users explicitly specify mori as all2all_backend, on vLLM side it does not construct another mori a2a manager. Effectively all all2all_backend specified at vLLM frontend will be swapped out by the lightweight AgRs, but sure I can use a recipe to articulate this.

[zejunchen-zejun]

Thank you for help enabling DBO.
Could you help:

• check the DBO perf gain with high concurrency, you can use GPTOSS model
• add deepseekv3.2 DBO into atom-vllm nightly accuracy check