Canlin Guo

@Bounty-hunter Could we make the final log not a ERROR when the offline script works normally? My concern is the red ERROR word will make users confused and misunderstood it...

Thanks for suggesting! Let's add the `soft_fail` in #346 first. BTW, the current error of NPU CI is seems because of not rebasing on the newest changes.

Also cc @tjtanaa.

vllm(vllm-ascend) and vllm-omni are different. The former only supports the omni-thinker, while the latter supports not only omni-thinker but also talker and code2wav to generate the audio. So even if...

You could refer to the [Dockerfile.ci.npu](https://github.com/vllm-project/vllm-omni/blob/main/docker/Dockerfile.ci.npu), which has been verified in CI. But if you want to experience qwen3-omni, you could add the below command: ``` git fetch origin pull/266/head:pr-266...

> Is it possible to degrade to forward_native for NPU operation when mindiesd is not installed? Agree. We should make mindie-sd an optional acceleration dependency on NPU.

The key error is missing the adaption of teacache for the new ops `AdaLayerNormer`. @jiangmengyu18 ``` [2025-12-26T14:09:08Z] [Stage-0] ERROR 12-26 06:09:08 [gpu_worker.py:222] File "/workdir/vllm_omni/diffusion/cache/teacache/extractors.py", line 211, in extract_qwen_context [2025-12-26T14:09:08Z] [Stage-0]...

happy to see it's ready for merging now :)

Here is the similar issue and temporary solution: https://github.com/vllm-project/vllm-omni/issues/208#issuecomment-3615829963. After upgrading to v0.12.0, this issue will be resolved.

> > Here is the similar issue and temporary solution: [#208 (comment)](https://github.com/vllm-project/vllm-omni/issues/208#issuecomment-3615829963). > > After upgrading to v0.12.0, this issue will be resolved. > > I switched to vllm-ascend v0.12.0.rc1,...