[NPU][Model] Support Qwen3-Omni on NPU

[gcanlin]

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

Purpose

Following #266, support Qwen3-Omni on NPU.

Test Plan

python end2end.py --output-wav output_audio \
                   --query-type text

Test Result

INFO 12-26 06:40:55 [log_utils.py:529] {'type': 'request_level_metrics',
INFO 12-26 06:40:55 [log_utils.py:529]  'request_id': '0_58ef1eab-5f96-4eaf-88cf-bf5d0fa306f8',
INFO 12-26 06:40:55 [log_utils.py:529]  'e2e_time_ms': 38717.89574623108,
INFO 12-26 06:40:55 [log_utils.py:529]  'e2e_tpt': 0.0,
INFO 12-26 06:40:55 [log_utils.py:529]  'num_tokens_out': 0,
INFO 12-26 06:40:55 [log_utils.py:529]  'transfers_total_time_ms': 9.086847305297852,
INFO 12-26 06:40:55 [log_utils.py:529]  'transfers_total_bytes': 2194264,
INFO 12-26 06:40:55 [log_utils.py:529]  'stages': {0: {'stage_gen_time_ms': 2264.1279697418213, 'num_tokens_out': 27},
INFO 12-26 06:40:55 [log_utils.py:529]             1: {'stage_gen_time_ms': 13507.445573806763, 'num_tokens_out': 110},
INFO 12-26 06:40:55 [log_utils.py:529]             2: {'stage_gen_time_ms': 21961.585521697998, 'num_tokens_out': 0}}}
INFO 12-26 06:40:55 [omni_llm.py:476] [Summary] {'e2e_requests': 1,
INFO 12-26 06:40:55 [omni_llm.py:476]  'e2e_total_time_ms': 38718.159437179565,
INFO 12-26 06:40:55 [omni_llm.py:476]  'e2e_sum_time_ms': 38717.89574623108,
INFO 12-26 06:40:55 [omni_llm.py:476]  'e2e_total_tokens': 0,
INFO 12-26 06:40:55 [omni_llm.py:476]  'e2e_avg_time_per_request_ms': 38717.89574623108,
INFO 12-26 06:40:55 [omni_llm.py:476]  'e2e_avg_tokens_per_s': 0.0,
INFO 12-26 06:40:55 [omni_llm.py:476]  'wall_time_ms': 38718.159437179565,
INFO 12-26 06:40:55 [omni_llm.py:476]  'final_stage_id': {'0_58ef1eab-5f96-4eaf-88cf-bf5d0fa306f8': 2},
INFO 12-26 06:40:55 [omni_llm.py:476]  'stages': [{'stage_id': 0,
INFO 12-26 06:40:55 [omni_llm.py:476]              'requests': 1,
INFO 12-26 06:40:55 [omni_llm.py:476]              'tokens': 27,
INFO 12-26 06:40:55 [omni_llm.py:476]              'total_time_ms': 2277.573823928833,
INFO 12-26 06:40:55 [omni_llm.py:476]              'avg_time_per_request_ms': 2277.573823928833,
INFO 12-26 06:40:55 [omni_llm.py:476]              'avg_tokens_per_s': 11.854720016681956},
INFO 12-26 06:40:55 [omni_llm.py:476]             {'stage_id': 1,
INFO 12-26 06:40:55 [omni_llm.py:476]              'requests': 1,
INFO 12-26 06:40:55 [omni_llm.py:476]              'tokens': 110,
INFO 12-26 06:40:55 [omni_llm.py:476]              'total_time_ms': 13517.81678199768,
INFO 12-26 06:40:55 [omni_llm.py:476]              'avg_time_per_request_ms': 13517.81678199768,
INFO 12-26 06:40:55 [omni_llm.py:476]              'avg_tokens_per_s': 8.137408708371623},
INFO 12-26 06:40:55 [omni_llm.py:476]             {'stage_id': 2,
INFO 12-26 06:40:55 [omni_llm.py:476]              'requests': 1,
INFO 12-26 06:40:55 [omni_llm.py:476]              'tokens': 0,
INFO 12-26 06:40:55 [omni_llm.py:476]              'total_time_ms': 21966.142654418945,
INFO 12-26 06:40:55 [omni_llm.py:476]              'avg_time_per_request_ms': 21966.142654418945,
INFO 12-26 06:40:55 [omni_llm.py:476]              'avg_tokens_per_s': 0.0}],
INFO 12-26 06:40:55 [omni_llm.py:476]  'transfers': [{'from_stage': 0,
INFO 12-26 06:40:55 [omni_llm.py:476]                 'to_stage': 1,
INFO 12-26 06:40:55 [omni_llm.py:476]                 'samples': 1,
INFO 12-26 06:40:55 [omni_llm.py:476]                 'total_bytes': 2188830,
INFO 12-26 06:40:55 [omni_llm.py:476]                 'total_time_ms': 3.2148361206054688,
INFO 12-26 06:40:55 [omni_llm.py:476]                 'tx_mbps': 5446.8219663719965,
INFO 12-26 06:40:55 [omni_llm.py:476]                 'rx_samples': 1,
INFO 12-26 06:40:55 [omni_llm.py:476]                 'rx_total_bytes': 2188830,
INFO 12-26 06:40:55 [omni_llm.py:476]                 'rx_total_time_ms': 3.0269622802734375,
INFO 12-26 06:40:55 [omni_llm.py:476]                 'rx_mbps': 5784.888736181474,
INFO 12-26 06:40:55 [omni_llm.py:476]                 'total_samples': 1,
INFO 12-26 06:40:55 [omni_llm.py:476]                 'total_transfer_time_ms': 6.963253021240234,
INFO 12-26 06:40:55 [omni_llm.py:476]                 'total_mbps': 2514.7212009367936},
INFO 12-26 06:40:55 [omni_llm.py:476]                {'from_stage': 1,
INFO 12-26 06:40:55 [omni_llm.py:476]                 'to_stage': 2,
INFO 12-26 06:40:55 [omni_llm.py:476]                 'samples': 1,
INFO 12-26 06:40:55 [omni_llm.py:476]                 'total_bytes': 5434,
INFO 12-26 06:40:55 [omni_llm.py:476]                 'total_time_ms': 0.2963542938232422,
INFO 12-26 06:40:55 [omni_llm.py:476]                 'tx_mbps': 146.68928679646018,
INFO 12-26 06:40:55 [omni_llm.py:476]                 'rx_samples': 1,
INFO 12-26 06:40:55 [omni_llm.py:476]                 'rx_total_bytes': 5434,
INFO 12-26 06:40:55 [omni_llm.py:476]                 'rx_total_time_ms': 1.2280941009521484,
INFO 12-26 06:40:55 [omni_llm.py:476]                 'rx_mbps': 35.397938941564746,
INFO 12-26 06:40:55 [omni_llm.py:476]                 'total_samples': 1,
INFO 12-26 06:40:55 [omni_llm.py:476]                 'total_transfer_time_ms': 2.123594284057617,
INFO 12-26 06:40:55 [omni_llm.py:476]                 'total_mbps': 20.470953574491972}]}

text:

Prompt:
<|im_start|>system
You are Qwen, a virtual human developed by the Qwen Team, Alibaba Group, capable of perceiving auditory and visual inputs, as well as generating text and speech.<|im_end|>
<|im_start|>user
Explain the system architecture for a scalable audio generation pipeline. Answer in 15 words.<|im_end|>
<|im_start|>assistant

vllm_text_output:
Modular microservices, distributed processing, load balancing, cloud scalability, real-time streaming, and GPU-accelerated inference.

audio(use ffmpeg to convert audio to mp4 format for showing):

ffmpeg -i output_audio/output_0_58ef1eab-5f96-4eaf-88cf-bf5d0fa306f8.wav -c:a aac -b:a 192k output.mp4

output-qwen3-omni.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)

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[gcanlin]

@hsliuustc0106 Could we merge it now? We want to ensure Qwen3-Omni basic functionality on NPU in the main at least. Thanks for reviewing again!

[hsliuustc0106]

lgtm