[BugFix] Fix NoneType' object has no attribute 'detach'

[amy-why-3459]

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

Purpose

fix #2788

k:tts_bos_embed, v:[tensor([[[ 0.0003, -0.0022,  0.0005,  ...,  0.0018,  0.0003,  0.0007]]],
(Worker pid=77251) INFO 04-14 12:29:51 [gpu_ar_model_runner.py:800]        dtype=torch.bfloat16)], idx:3, len(v):1

When the key is something like tts_bos_embed, idx will be greater than len(v), causing the key to not be assigned a value, thus resulting in None when detaching.

Test Plan

vllm bench serve \
    --omni \
  --dataset-name random \
  --port 28889 \
  --max-concurrency 4 \
  --model /home/models/Qwen/Qwen3-Omni-30B-A3B-Instruct \
  --endpoint /v1/chat/completions \
  --backend openai-chat-omni \
  --num-prompts 8 \
  --random-input-len 2500 \
  --ignore-eos \
  --percentile-metrics ttft,tpot,itl,e2el,audio_ttfp,audio_rtf \
  --random-output-len 900 \
  --extra_body '{"modalities": ["text", "audio"]}'

Test Result

============ Serving Benchmark Result ============
Successful requests:                     8
Failed requests:                         0
Maximum request concurrency:             4
Benchmark duration (s):                  199.45
Request throughput (req/s):              0.04
Peak concurrent requests:                5.00
----------------End-to-end Latency----------------
Mean E2EL (ms):                          87468.74
Median E2EL (ms):                        87588.02
P99 E2EL (ms):                           106065.05
================== Text Result ===================
Total input tokens:                      20112
Total generated tokens:                  6304
Output token throughput (tok/s):         31.61
Peak output token throughput (tok/s):    206.00
Peak concurrent requests:                5.00
Total Token throughput (tok/s):          132.45
---------------Time to First Token----------------
Mean TTFT (ms):                          1393.74
Median TTFT (ms):                        512.06
P99 TTFT (ms):                           3004.43
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          27.19
Median TPOT (ms):                        22.15
P99 TPOT (ms):                           52.22
---------------Inter-token Latency----------------
Mean ITL (ms):                           21.44
Median ITL (ms):                         0.01
P99 ITL (ms):                            330.02
================== Audio Result ==================
Total audio duration generated(s):       2153.29
Total audio frames generated:            51678825
Audio throughput(audio duration/s):      10.80
---------------Time to First Packet---------------
Mean AUDIO_TTFP (ms):                    3382.07
Median AUDIO_TTFP (ms):                  3047.07
P99 AUDIO_TTFP (ms):                     5916.30
-----------------Real Time Factor-----------------
Mean AUDIO_RTF:                          0.33
Median AUDIO_RTF:                        0.33
P99 AUDIO_RTF:                           0.34
==================================================

---

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. Please provide the test scripts & test commands. Please state the reasons if your codes don't require additional test scripts. For test file guidelines, please check the test style doc
• The test results. Please paste the results comparison before and after, or the e2e results.
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model. Please run mkdocs serve to sync the documentation editions to ./docs.
• (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 usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.
Credits must be used to enable repository wide code reviews.

[amy-why-3459]

@Sy0307 @linyueqian @yenuo26 PTAL

[linyueqian]

LGTM

[gcanlin]

Could you also sync the code to NPUARModelRunner please?

[amy-why-3459]

This was modified in this PR #2690, and it was found that npu_ar_model_runner.py was not modified, so NPU does not need to be modified.

element = v[idx] if idx < len(v) else v[0]
# Clone tensors to avoid cross-request aliasing
if isinstance(element, torch.Tensor):
    element = element.clone()
mm_payload[k] = element

[amy-why-3459]

@linyueqian @Gaohan123 @gcanlin @hsliuustc0106 I think this PR is ready and can be merged. Ideally, it should be merged tonight, otherwise it will affect the nightly test results.