[Refactor] Reuse upstream Qwen3MoeSparseMoeBlock

[gcanlin]

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

Purpose

After vLLM-Omni upgrade vLLM to v0.15.0, we can reuse vllm-project/vllm#32082 and remove the redundant talker MoE safely.

Test Plan

bash run_single_prompt.sh

Test Result

O 02-04 15:27:33 [log_utils.py:550] {'type': 'request_level_metrics',
INFO 02-04 15:27:33 [log_utils.py:550]  'request_id': '0_93100e0c-8a35-4d81-8927-feb32c55b762',
INFO 02-04 15:27:33 [log_utils.py:550]  'e2e_time_ms': 32853.352546691895,
INFO 02-04 15:27:33 [log_utils.py:550]  'e2e_tpt': 98.65871635643211,
INFO 02-04 15:27:33 [log_utils.py:550]  'e2e_total_tokens': 333,
INFO 02-04 15:27:33 [log_utils.py:550]  'transfers_total_time_ms': 28.17392349243164,
INFO 02-04 15:27:33 [log_utils.py:550]  'transfers_total_bytes': 6988121,
INFO 02-04 15:27:33 [log_utils.py:550]  'stages': {0: {'stage_gen_time_ms': 7010.502099990845,
INFO 02-04 15:27:33 [log_utils.py:550]                 'num_tokens_out': 15,
INFO 02-04 15:27:33 [log_utils.py:550]                 'num_tokens_in': 265},
INFO 02-04 15:27:33 [log_utils.py:550]             1: {'stage_gen_time_ms': 5290.610313415527, 'num_tokens_out': 53},
INFO 02-04 15:27:33 [log_utils.py:550]             2: {'stage_gen_time_ms': 93.02067756652832, 'num_tokens_out': 0}}}
Processed prompts: 100%|██████████████████████████████████| 1/1 [00:32<00:00, 32.85s/req, est. speed stage-2 tok/s: 10.14, avg e2e_lat: 0.0ms]
INFO 02-04 15:27:33 [omni.py:860] [Summary] {'e2e_requests': 1, [00:32<00:00, 32.85s/req, est. speed stage-2 tok/s: 10.14, avg e2e_lat: 0.0ms]
INFO 02-04 15:27:33 [omni.py:860]  'e2e_total_time_ms': 32854.35175895691,
INFO 02-04 15:27:33 [omni.py:860]  'e2e_sum_time_ms': 32853.352546691895,
INFO 02-04 15:27:33 [omni.py:860]  'e2e_total_tokens': 333,
INFO 02-04 15:27:33 [omni.py:860]  'e2e_avg_time_per_request_ms': 32853.352546691895,
INFO 02-04 15:27:33 [omni.py:860]  'e2e_avg_tokens_per_s': 10.135951864478159,
INFO 02-04 15:27:33 [omni.py:860]  'wall_time_ms': 32854.35175895691,
INFO 02-04 15:27:33 [omni.py:860]  'final_stage_id': {'0_93100e0c-8a35-4d81-8927-feb32c55b762': 2},
INFO 02-04 15:27:33 [omni.py:860]  'stages': [{'stage_id': 0,
INFO 02-04 15:27:33 [omni.py:860]              'requests': 1,
INFO 02-04 15:27:33 [omni.py:860]              'tokens': 280,
INFO 02-04 15:27:33 [omni.py:860]              'total_time_ms': 17102.519512176514,
INFO 02-04 15:27:33 [omni.py:860]              'avg_time_per_request_ms': 17102.519512176514,
INFO 02-04 15:27:33 [omni.py:860]              'avg_tokens_per_s': 16.37185677821609},
INFO 02-04 15:27:33 [omni.py:860]             {'stage_id': 1,
INFO 02-04 15:27:33 [omni.py:860]              'requests': 1,
INFO 02-04 15:27:33 [omni.py:860]              'tokens': 53,
INFO 02-04 15:27:33 [omni.py:860]              'total_time_ms': 15376.47795677185,
INFO 02-04 15:27:33 [omni.py:860]              'avg_time_per_request_ms': 15376.47795677185,
INFO 02-04 15:27:33 [omni.py:860]              'avg_tokens_per_s': 3.4468231378472876},
INFO 02-04 15:27:33 [omni.py:860]             {'stage_id': 2,
INFO 02-04 15:27:33 [omni.py:860]              'requests': 1,
INFO 02-04 15:27:33 [omni.py:860]              'tokens': 0,
INFO 02-04 15:27:33 [omni.py:860]              'total_time_ms': 102.47135162353516,
INFO 02-04 15:27:33 [omni.py:860]              'avg_time_per_request_ms': 102.47135162353516,
INFO 02-04 15:27:33 [omni.py:860]              'avg_tokens_per_s': 0.0}],
INFO 02-04 15:27:33 [omni.py:860]  'transfers': [{'from_stage': 0,
INFO 02-04 15:27:33 [omni.py:860]                 'to_stage': 1,
INFO 02-04 15:27:33 [omni.py:860]                 'samples': 1,
INFO 02-04 15:27:33 [omni.py:860]                 'total_bytes': 5964365,
INFO 02-04 15:27:33 [omni.py:860]                 'total_time_ms': 12.66622543334961,
INFO 02-04 15:27:33 [omni.py:860]                 'tx_mbps': 3767.0985923216504,
INFO 02-04 15:27:33 [omni.py:860]                 'rx_samples': 1,
INFO 02-04 15:27:33 [omni.py:860]                 'rx_total_bytes': 5964365,
INFO 02-04 15:27:33 [omni.py:860]                 'rx_total_time_ms': 10.818004608154297,
INFO 02-04 15:27:33 [omni.py:860]                 'rx_mbps': 4410.695107675761,
INFO 02-04 15:27:33 [omni.py:860]                 'total_samples': 1,
INFO 02-04 15:27:33 [omni.py:860]                 'total_transfer_time_ms': 24.20973777770996,
INFO 02-04 15:27:33 [omni.py:860]                 'total_mbps': 1970.8978444174388},
INFO 02-04 15:27:33 [omni.py:860]                {'from_stage': 1,
INFO 02-04 15:27:33 [omni.py:860]                 'to_stage': 2,
INFO 02-04 15:27:33 [omni.py:860]                 'samples': 1,
INFO 02-04 15:27:33 [omni.py:860]                 'total_bytes': 1023756,
INFO 02-04 15:27:33 [omni.py:860]                 'total_time_ms': 1.355886459350586,
INFO 02-04 15:27:33 [omni.py:860]                 'tx_mbps': 6040.364179108845,
INFO 02-04 15:27:33 [omni.py:860]                 'rx_samples': 1,
INFO 02-04 15:27:33 [omni.py:860]                 'rx_total_bytes': 1023756,
INFO 02-04 15:27:33 [omni.py:860]                 'rx_total_time_ms': 1.9745826721191406,
INFO 02-04 15:27:33 [omni.py:860]                 'rx_mbps': 4147.736185292441,
INFO 02-04 15:27:33 [omni.py:860]                 'total_samples': 1,
INFO 02-04 15:27:33 [omni.py:860]                 'total_transfer_time_ms': 3.9641857147216797,
INFO 02-04 15:27:33 [omni.py:860]                 'total_mbps': 2066.010169398689}]}
Adding requests:   0%|                                                                                                  | 0/1 [00:32<?, ?it/s]
query type: use_audio
Request ID: 0_93100e0c-8a35-4d81-8927-feb32c55b762, Text saved to output_audio/0_93100e0c-8a35-4d81-8927-feb32c55b762.txt
Request ID: 0_93100e0c-8a35-4d81-8927-feb32c55b762, Saved audio to output_audio/output_0_93100e0c-8a35-4d81-8927-feb32c55b762.wav

output-2.mp4

---

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)

[gcanlin]

cc @Isotr0py

[hsliuustc0106]

hiw about he acc?

[gcanlin]

hiw about he acc?

I uploaded the audio in the description. The accuracy is correct. Before this PR, we used monkey patch way to add the SharedFusedMoE. Removing it would make code clean and healthy.