[RFC]: vLLM-Omni NPU 2026 Q1 Roadmap

[gcanlin]

Background

We have completed the initial Ascend NPU enablement in vllm-omni v0.11.0rc1 and v0.12.0rc1, with support for most mainstream models such as Qwen3-Omni and the Qwen-Image series.

Building on this foundation, the next phase will focus on systematically expanding model coverage and prioritizing performance optimization efforts, with a clear roadmap to improve scalability, stability, and overall serving efficiency on Ascend NPU.

Version match

Currently, vLLM-Omni’s NPU support depends on vLLM-Ascend, the Ascend support plugin of vLLM. The AR (auto-regressive) path is jointly supported by vLLM and vLLM-Ascend.

Meanwhile, MindIE-SD serves as a standalone Ascend-optimized diffusion operator library. It is currently integrated through the FlashAttentionBackend and a set of CustomOp, delivering Ascend-native operators to improve the performance of diffusion models.

We're also building the separate plugin platform in vLLM-Omni to support scalable hardware better in the future.

vLLM	vLLM-Ascend	vLLM-Omni	MindIE-SD(Optional)	status
v0.11.0	v0.11.0rc2	v0.11.0rc1	NA	released
v0.12.0	v0.12.0rc1	v0.12.0rc1	main	released
v0.14.0	v0.14.0rc1	v0.14.0	main	released
v0.15.0	v0.15.0rc1	v0.15.0rc1	main	skipped
v0.16.0	e2175d9	v0.16.0	main	released
v0.17.0	v0.17.0rc1	9718a9 ~8e120	main	released
v0.18.0	v0.18.0rc1	v0.18.0	main	released
v0.19.0	f1f0870	main	main	developing

How to install MindIE-SD

Official Link: MindIE-SD

We are actively working to simplify the installation of mindie-sd. Eventually, it will be available via pip install mindie-sd. At the moment, however, some additional work is required.

git clone https://gitcode.com/Ascend/MindIE-SD.git && cd MindIE-SD
# Need to comment the line `source ${current_script_dir}/build_tik_ops.sh` in build/build_ops.sh
sed -i 's|^\(\s*\)source ${current_script_dir}/build_tik_ops.sh|\1# source ${current_script_dir}/build_tik_ops.sh|' build/build_ops.sh
python setup.py bdist_wheel
cd dist
pip install mindiesd-*.whl

Feature Support

Omni(AR+Generator) Pipeline

• Async chunk: follow this RFC [RFC]: Support async computation and communication across stages by chunks #268
• TTS Performance Optimization: [RFC]: TTS Model Performance Optimization on NPU #1600
• Support code2wav multi-batch
• Streaming input and output
• Talker ACL graph Support
• More Ascend-friendly Ops
• Expert Parallelism (EP)

Diffusion Pipeline

• Support sparse attention backend by integrating the sparse attention interface from MindIE-SD
• Support LA from MindIE-SD
• Following [RFC]: Diffusion Models Features Supports Plan #814's features, make sure them work on NPU
• Qwen-Image-Edit-2511 Optimization
• Wan2.2 Optimization: [RFC]: Wan2.2 Performance Optimization Roadmap on vLLM-Omni  #1355
• Remove NPU hardcode: [Feat] support SP for FLUX.2-klein #1250 (comment)
• Refactor ring attention for hardware dispatch: [Bugfix] Fix ring attention for npu device #755

Others(UX & Hardware Scalable)

• Platform: [Hardware] Support platforms and plugin system #774
• Disable torch compile by default: [Platform] Add supports_torch_inductor interface #1108
• Dependencies router: [Hardware] [Feat] Setup platform dependent package installation #1046

Docs

Known Issues

• Memory usage: currently, Qwen2.5-Omni and Qwen3-Omni have to separate talker to one different device from thinker. We expect to make them together so that Qwen2.5-Omni and Qwen3-Omni would only need 2 and 4 cards.
• Qwen2.5-Omni: enabling ACL graph leads to accuracy problem. [Bug]: vllm-omni(v0.12.0) results of talker model of qwen2.5-omni are incorrect when running with enforce eager being False #912
• Qwen3-Omni: talker ACL graph breaks. [NPU] Align with GPUModelRunner #1114 (comment)
• [Feature][NPU]: Under the same request, the quality of the raw image on the NPU differs from that on the NVIDIA GPU. #1322: If you obtain images on NPU that differ from those on GPU, this is normal and expected, as long as there is no obvious degradation (e.g., blurred faces, failure to follow the prompt, etc.). You can make the outputs closer by fixing the same seed and using a CPU-based generator(online server can use [Misc] Add per-request generator_device to online image gen and edit #1183). However, even in that case, minor differences may still remain. The root cause is that different hardware backends cannot perfectly align in PyTorch operator implementations, such as conv3d, nn.Linear, and others.

Model Support List

Architecture	Models	Example HF Models	NPU support
Qwen3OmniMoeForConditionalGeneration	Qwen3-Omni	Qwen/Qwen3-Omni-30B-A3B-Instruct	✅
Qwen2_5OmniForConditionalGeneration	Qwen2.5-Omni	Qwen/Qwen2.5-Omni-7B, Qwen/Qwen2.5-Omni-3B	✅
BagelForConditionalGeneration	BAGEL (DiT-only)	ByteDance-Seed/BAGEL-7B-MoT
QwenImagePipeline	Qwen-Image	Qwen/Qwen-Image	✅
QwenImagePipeline	Qwen-Image-2512	Qwen/Qwen-Image-2512	✅
QwenImageEditPipeline	Qwen-Image-Edit	Qwen/Qwen-Image-Edit	✅
QwenImageEditPlusPipeline	Qwen-Image-Edit-2509	Qwen/Qwen-Image-Edit-2509	✅
QwenImageLayeredPipeline	Qwen-Image-Layered	Qwen/Qwen-Image-Layered	✅
ZImagePipeline	Z-Image	Tongyi-MAI/Z-Image-Turbo	✅
WanPipeline	Wan2.2-T2V, Wan2.2-TI2V	Wan-AI/Wan2.2-T2V-A14B-Diffusers, Wan-AI/Wan2.2-TI2V-5B-Diffusers	✅
WanImageToVideoPipeline	Wan2.2-I2V	Wan-AI/Wan2.2-I2V-A14B-Diffusers	✅
OvisImagePipeline	Ovis-Image	OvisAI/Ovis-Image
LongcatImagePipeline	LongCat-Image	meituan-longcat/LongCat-Image	✅
LongCatImageEditPipeline	LongCat-Image-Edit	meituan-longcat/LongCat-Image-Edit
StableDiffusion3Pipeline	Stable-Diffusion-3	stabilityai/stable-diffusion-3.5-medium
Flux2KleinPipeline	FLUX.2-klein	black-forest-labs/FLUX.2-klein-4B, black-forest-labs/FLUX.2-klein-9B
StableAudioPipeline	Stable-Audio-Open	stabilityai/stable-audio-open-1.0
Qwen3TTSForConditionalGeneration	Qwen3-TTS-12Hz-1.7B-CustomVoice	Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice	✅
Qwen3TTSForConditionalGeneration	Qwen3-TTS-12Hz-1.7B-VoiceDesign	Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign	✅
Qwen3TTSForConditionalGeneration	Qwen3-TTS-12Hz-1.7B-Base	Qwen/Qwen3-TTS-12Hz-0.6B-Base	✅

Feedback Period.

No response

CC List.

No response

Any Other Things.

No response

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.

[hsliuustc0106]

I think you can also add the installation guidance here: how it can work together with vllm-ascend and MindIE-SD acceleration on NPU

[gcanlin]

I think you can also add the installation guidance here: how it can work together with vllm-ascend and MindIE-SD acceleration on NPU

Good idea :) Let me add it later.

[yjb767868009]

I noticed that the ti2v and i2v pipelines for Wan2.2 on NPU haven't been adapted yet. Could I take on the development of this requirement?

[gcanlin]

@yjb767868009 Of course. Thanks for your contribution!

[scalerinfo]

@gcanlin Are the models listed in the Model Support List already supported in the current version (e.g. v0.12.0rc1)?

[gcanlin]

@gcanlin Are the models listed in the Model Support List already supported in the current version (e.g. v0.12.0rc1)?

v0.12.0rc1 only supports Qwen-Omni series and Qwen-Image series. It's better to use the main branch with vLLM 0.14.0 and vLLM-Ascend v0.14.0rc1. You could experience more models like Qwen3-TTS.

[scalerinfo]

@gcanlin OK, Thanks for your reply