[Core] add clean up method for diffusion engine

[ZJY0516]

Purpose

• add clean up method for diffusion engine
• add a test for qwen-image

Test Plan

Test Result

---

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]

The account who enabled Codex for this repo no longer has access to Codex. Please contact the admins of this repo to enable Codex again.

[ZJY0516]

cc @gcanlin This pr changes worker

[hsliuustc0106]

shall we move this omni_diffusion file under entrypoints since omni_llm is also listed there?

[ZJY0516]

shall we move this omni_diffusion file under entrypoints since omni_llm is also listed there?

Yes

[SamitHuang]

will these semaphore leakage warning resolved by this PR?

/usr/lib/python3.12/multiprocessing/resource_tracker.py:254: UserWarning: resource_tracker: There appear to be 1 leaked semaphore objects to clean up at shutdown
  warnings.warn('resource_tracker: There appear to be %d '
/usr/lib/python3.12/multiprocessing/resource_tracker.py:254: UserWarning: resource_tracker: There appear to be 2 leaked shared_memory objects to clean up at shutdown
  warnings.warn('resource_tracker: There appear to be %d '

[ZJY0516]

will these semaphore leakage warning resolved by this PR?

/usr/lib/python3.12/multiprocessing/resource_tracker.py:254: UserWarning: resource_tracker: There appear to be 1 leaked semaphore objects to clean up at shutdown
  warnings.warn('resource_tracker: There appear to be %d '
/usr/lib/python3.12/multiprocessing/resource_tracker.py:254: UserWarning: resource_tracker: There appear to be 2 leaked shared_memory objects to clean up at shutdown
  warnings.warn('resource_tracker: There appear to be %d '

Yes. But there is a new warning now. I'll resolve it later

[rank0]:[W1206 13:49:30.182850024 ProcessGroupNCCL.cpp:1538] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())

[ZJY0516]

will these semaphore leakage warning resolved by this PR?

/usr/lib/python3.12/multiprocessing/resource_tracker.py:254: UserWarning: resource_tracker: There appear to be 1 leaked semaphore objects to clean up at shutdown
  warnings.warn('resource_tracker: There appear to be %d '
/usr/lib/python3.12/multiprocessing/resource_tracker.py:254: UserWarning: resource_tracker: There appear to be 2 leaked shared_memory objects to clean up at shutdown
  warnings.warn('resource_tracker: There appear to be %d '

Yes. But there is a new warning now. I'll resolve it later

[rank0]:[W1206 13:49:30.182850024 ProcessGroupNCCL.cpp:1538] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())

Done

[SamitHuang]

LGTM

[gcanlin]

Qwen-Image on NPU can work normally. Thanks!

INFO 12-08 03:16:35 [utils.py:474] No adjustment needed for ACL graph batch sizes: Qwen3ForCausalLM model (layers: 28) with 35 sizes
WARNING 12-08 03:16:35 [platform.py:275] If chunked prefill or prefix caching is enabled, block size must be set to 128.
WARNING 12-08 03:16:35 [__init__.py:755] Current vLLM config is not set.
WARNING 12-08 03:16:35 [platform.py:143] Model config is missing. This may indicate that we are running a test case
WARNING 12-08 03:16:35 [platform.py:275] If chunked prefill or prefix caching is enabled, block size must be set to 128.
WARNING 12-08 03:16:35 [parallel_state.py:1055] Distributed backend nccl is not available; falling back to gloo.
WARNING 12-08 03:16:35 [__init__.py:755] Current vLLM config is not set.
WARNING 12-08 03:16:35 [platform.py:143] Model config is missing. This may indicate that we are running a test case
WARNING 12-08 03:16:35 [platform.py:275] If chunked prefill or prefix caching is enabled, block size must be set to 128.
INFO 12-08 03:16:35 [parallel_state.py:1208] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, TP rank 0, EP rank 0
WARNING:vllm_omni.diffusion.attention.backends.flash_attn:FlashAttentionBackend is not available. You may install flash-attn by running `uv pip install flash-attn==2.8.1 --no-build-isolation` or install pre-built flash-attn from https://github.com/Dao-AILab/flash-attention/releases
Loading checkpoint shards: 100%|████████████████████████████████████████████████████████████████████████████████| 4/4 [00:02<00:00,  1.38it/s]
/usr/local/python3.11.13/lib/python3.11/site-packages/torch/utils/_device.py:104: UserWarning: Cannot create tensor with interal format while allow_internel_format=False, tensor will be created with base format. (Triggered internally at build/CMakeFiles/torch_npu.dir/compiler_depend.ts:335.)
  return func(*args, **kwargs)
WARNING 12-08 03:16:41 [__init__.py:755] Current vLLM config is not set.
WARNING 12-08 03:16:41 [platform.py:143] Model config is missing. This may indicate that we are running a test case
WARNING 12-08 03:16:41 [platform.py:275] If chunked prefill or prefix caching is enabled, block size must be set to 128.
INFO:vllm_omni.diffusion.worker.npu.npu_worker:Worker 0: Initialized device, model, and distributed environment.
INFO:vllm_omni.diffusion.worker.npu.npu_worker:Worker 0: Model loaded successfully.
INFO:vllm_omni.diffusion.worker.npu.npu_worker:Worker 0: Scheduler loop started.
INFO:vllm_omni.diffusion.worker.npu.npu_worker:Worker 0 ready to receive requests via shared memory
INFO:vllm_omni.diffusion.scheduler:SyncScheduler initialized result MessageQueue
INFO:vllm_omni.diffusion.omni_diffusion:Prepared 1 requests for generation.
('Warning: torch.save with "_use_new_zipfile_serialization = False" is not recommended for npu tensor, which may bring unexpected errors and hopefully set "_use_new_zipfile_serialization = True"', 'if it is necessary to use this, please convert the npu tensor to cpu tensor for saving')
('Warning: torch.save with "_use_new_zipfile_serialization = False" is not recommended for npu tensor, which may bring unexpected errors and hopefully set "_use_new_zipfile_serialization = True"', 'if it is necessary to use this, please convert the npu tensor to cpu tensor for saving')
Warning: The current version of the file storing weights is old, and it is relanded due to internal bug of torch and compatibility issue. We will deprecate the loading support for this type of file in the future, please use newer torch to re-store the weight file.
INFO:vllm_omni.diffusion.diffusion_engine:Generation completed successfully.
Saved generated image to outputs/coffee.png
INFO:vllm_omni.diffusion.worker.npu.npu_worker:Worker 0: Received shutdown message
INFO:vllm_omni.diffusion.worker.npu.npu_worker:event loop terminated.
INFO:vllm_omni.diffusion.worker.npu.npu_worker:Worker 0: Destroyed process group
INFO:vllm_omni.diffusion.worker.npu.npu_worker:Worker 0: Shutdown complete.

[hsliuustc0106]

why omni_diffusion.py is not moved to entrypoints?

[hsliuustc0106]

@ZJY0516

[ZJY0516]

why omni_diffusion.py is not moved to entrypoints?

I'll do it later