[Bug]: vllm-omni(v0.12.0) results of talker model of qwen2.5-omni are incorrect when running with enforce eager being False

[Jinxinxiang5525]

Your current environment

The output of python collect_env.py

Your output of `python collect_env.py` here

Your code version

The commit id or version of vllm

The commit id or version of vllm-omni

🐛 Describe the bug

The tokens output by the talker model of qwen2.5-omni are incorrect when enforce eager (in qwen2_5_omni.yaml) is set to false (running compile on NPU), but it's correct for thinker model under the same parameters

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

[Jinxinxiang5525]

No errors when running, but the tokens output by talker model are incorrect.

[gcanlin]

Thanks for reporting. I will take a look after upgrading to v0.14.0.

[hsliuustc0106]

Thanks for reporting. I will take a look after upgrading to v0.14.0.

is aclgraph support for NPU?

[gcanlin]

Thanks for reporting. I will take a look after upgrading to v0.14.0.

is aclgraph support for NPU?

Qwen2.5-Omni should be supported in v0.12.0rc1 from my local test. I didn't pay more attention to acc before. So it needs to have a double check in the new version.