[NPU] Upgrade to v0.14.0

[gcanlin]

PLEASE FILL IN THE PR DESCRIPTION HERE ENSURING ALL CHECKLIST ITEMS (AT THE BOTTOM) HAVE BEEN CONSIDERED.

Purpose

This PR:

• Upgrade to v0.14.0. Align with vLLM v0.14.0 and vLLM-Ascend v0.14.0rc1.
• Align with GPU model runner to support MTP ACL graph.
• Remove the temporary patch for vLLM-Ascend.

Test Plan

NPU:

• Qwen3-Omni
  • eager
  • ACL graph
• Qwen2.5-Omni
  • eager
  • ACL graph(Will fix acc problem in a following-up PR)
• Qwen-Image
• Qwen3-TTS

GPU:

• CI

Test Result

Model	enforce_eager	time	example
Qwen3-Omni	true	81s	use_audio
Qwen3-Omni	false	70s	use_audio
Qwen2.5-Omni	true	180s	use_mixed_modalities
Qwen2.5-Omni	false	have acc problem(Will fix it in a following-up PR)	use_mixed_modalities

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft.

BEFORE SUBMITTING, PLEASE READ https://github.com/vllm-project/vllm-omni/blob/main/CONTRIBUTING.md (anything written below this line will be removed by GitHub Actions)

[hsliuustc0106]

paste the comparison gain in performance

[hsliuustc0106]

also, please help optimize the log for metrics? currently it's very hard to understand

[gcanlin]

@hsliuustc0106 I updated the benchmark result with #780 to compare before and after. But it's wired that enabling ACL graph makes performance regression... It needs more investigation. I will revert to eager mode by default. But this PR also fix the main branch breaking on NPU. So it's better to get it merged asap. Thanks!

also, please help optimize the log for metrics? currently it's very hard to understand

I think after benchmark is done, metrics seems to be unnecessary.

[hsliuustc0106]

could you fix this PR and test the result again? we expect this PR merged by next week

[gcanlin]

could you fix this PR and test the result again? we expect this PR merged by next week

After vLLM-Ascend v0.14.0rc1 releases, I will also upgrade NPUModelRunner to align v0.14.0rc1 in this PR.

[gcanlin]

@hsliuustc0106 @david6666666 Could you please take a look? This PR is ready. Docs and other things will be done in #671. Thanks!

[Gaohan123]

Why do we need the modification in the model file?

[gcanlin]

We add those properties so the wrapper presents the same interface as a normal Qwen MLP. vLLM-Ascend’s SharedFusedMoE validation splits the shared expert into gate_up_proj → act_fn → down_proj and expects these attributes to exist. The wrapper only holds the real MLP, so we forward them to keep validation and split execution working on NPU. It doesn't affect GPU behaviour. When we upgrade to vLLM v0.15.0 or later, we will remove this hack wrapper because upstream has implemented shared expert in Qwen3-MoE that we can reuse directly.