support fused_moe_allgather_ep

[lyj-jjj]

What this PR does / why we need it?

support fused_moe_allgather_ep

Does this PR introduce any user-facing change?

~~

How was this patch tested?

~~

[Yikun]

• Please update commits msg (epscially How was this patch tested?, should mention a E2E step how to reproduce)
• Add UT or E2E test to avoid this being break again after merge.

[Copilot]

Pull Request Overview

This PR adds support for a new fused MoE allgather endpoint path (fused_moe_allgather_ep) in both the dynamic quantization workflow and the DeepSeek model.

• Introduces fused_experts_with_allgather operator in w8a8_dynamic.py
• Registers and checks a new VLLM_ENABLE_FUSED_EXPERTS_ALLGATHER_EP env var and etp_group
• Updates DeepSeek V2 to conditionally prefer allgather-based MoE routing over all-to-all

Reviewed Changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 3 comments.

File	Description
vllm_ascend/quantization/w8a8_dynamic.py	Added fused_experts_with_allgather, imported get_etp_group
vllm_ascend/models/deepseek_v2.py	Added fused_experts_allgather_ep_enabled flag and logic tweaks
vllm_ascend/envs.py	Defined VLLM_ENABLE_FUSED_EXPERTS_ALLGATHER_EP environment var

Comments suppressed due to low confidence (2)

vllm_ascend/quantization/w8a8_dynamic.py:349

• This new public function lacks a docstring explaining its purpose, parameters, and return value. Please add a clear docstring to improve maintainability and usability.

def fused_experts_with_allgather(hidden_states: torch.Tensor,

vllm_ascend/models/deepseek_v2.py:332

• [nitpick] The flag enable_alltoall_ep is overloaded to control both all-to-all and allgather behaviors. Consider renaming it to something like alltoall_ep_enabled for clarity.

        else:

[lyj-jjj]

• Please update commits msg (epscially How was this patch tested?, should mention a E2E step how to reproduce)
• Add UT or E2E test to avoid this being break again after merge.

• Please update commits msg (epscially How was this patch tested?, should mention a E2E step how to reproduce)
• Add UT or E2E test to avoid this being break again after merge.

• Please update commits msg (epscially How was this patch tested?, should mention a E2E step how to reproduce)
• Add UT or E2E test to avoid this being break again after merge.

[jianzs]

This pr can be merged once the CI is happy. cc @Yikun @wangxiyuan

[wangxiyuan]

This PR is fine. But I'd like to merge this after #1229. what do you think? @jianzs @Yikun

[lyj-jjj]

I think my PR can be merged first because this PR is already ready and thoroughly tested with complete unit tests. It has been waiting to be merged for several days. The other PR involves significant refactoring and it's unclear when it can be merged, so I do not want my PR to be blocked by the other one. @wangxiyuan @Yikun @jianzs

[github-actions]

This pull request has conflicts, please resolve those before we can evaluate the pull request.