[Good First Issue]: Add vendor-organized community recipes for “run model X on hardware Y for task Z” in vLLM-Omni

Community Recipes To-Do List

All supported hardware platforms (CUDA, ROCm, NPU, XPU) are welcome for every entry.

Legend: 🙋 help wanted | ⏳ PR raised | ✅ merged

Vendor	Model	Task	Status	Claimer	PR	Notes
Omni-Modality
Qwen	Qwen3-Omni	omni chat / serving	✅			30B MoE (3B active)
Qwen	Qwen2.5-Omni	omni chat / speech	⏳	@allgather	#3151	7B / 3B
ByteDance	BAGEL-7B-MoT	image gen + understanding	⏳	@GrayMiao123		MoT, ~42 GiB VRAM
ByteDance	MammothModa2-Preview	text-to-image	🙋			AR + DiT pipeline
inclusionAI	Ming-flash-omni-2.0	omni-modal understanding	✅	@yuanheng-zhao	#2890	MoE, 4-GPU TP
SNU AIDAS	Dynin-Omni	omni-modal (t2t/i2t/s2t/t2i/v2t/t2s)	🙋			3-stage token-based
SenseNova	SenseNova-U1-8B-MoT	omni-modal (T2I/I2I/I2T/T2T)	✅	@princepride	#3319	MoE, ~36 GiB; think mode
Text-to-Image / Image Gen
Qwen	Qwen-Image	text-to-image	✅	@Semmer2	#2729	1024x1024, ~60 GiB
Qwen	Qwen-Image-2512	text-to-image	🙋	@AbelSara		updated variant
Qwen	Qwen-Image-Edit	image editing	⏳	@yixiaoer		text-guided editing
Qwen	Qwen-Image-Layered	layered image gen	🙋
Zhipu AI (GLM)	GLM-Image	image gen / editing	⏳	@friden-zhang		2-stage AR+DiT
Tongyi	Z-Image-Turbo	text-to-image (distilled)	⏳	@MrDongsls	#3283	4-9 steps, ~25 GiB; 2x RTX 5880 48GB
Stepfun	NextStep-1.1	text-to-image	⏳	@GrayMiao123		dual-level CFG, 512x512
Meituan	LongCat-Image	text-to-image	🙋	@Mashirona		1024x1024
Meituan	LongCat-Image-Edit	image editing	🙋
OvisAI	Ovis-Image	text-to-image	🙋			1024x1024
OmniGen2	OmniGen2	text-to-image	🙋	@Jerry2423 @Yuyi-Ao @jackywangno007-cyber		~20 GiB
Stability AI	Stable-Diffusion-3.5	text-to-image	🙋	@yangyonggit		1024x1024
Black Forest Labs	FLUX.1-dev	text-to-image	🙋			~78 GiB
Black Forest Labs	FLUX.1-schnell	text-to-image (fast)	🙋			fast distilled variant
Black Forest Labs	FLUX.2-dev	text-to-image	🙋			needs CPU offload on 80GiB
Black Forest Labs	FLUX.2-klein-4B	text-to-image	🙋			distilled, 4B params
Black Forest Labs	FLUX.2-klein-9B	text-to-image	🙋			distilled, 9B params
Tencent	HunyuanImage-3.0	image gen / understanding	✅	@Bounty-hunter	#2495	MoE, dual task
Baidu	ERNIE-Image	text-to-image	✅	@RuixiangMa	#2861	1024x1024, ~23 GiB
Text-to-Video / Image-to-Video
Tencent-Hunyuan	HunyuanVideo-1.5-T2V	text-to-video	⏳	@allgather	#3152	480p path on 1xA100; needs FP8 + VAE tiling for 720p
Tencent-Hunyuan	HunyuanVideo-1.5-I2V	image-to-video	⏳	@xldeng-chn		1xA100 80GB
Wan-AI	Wan2.2-T2V-A14B	text-to-video	⏳	@lengrongfu	#3018	720x1280, 81 frames, ~60 GiB
Wan-AI	Wan2.2-TI2V-5B	text+image-to-video	🙋			unified, 5B params
Wan-AI	Wan2.2-I2V-A14B	image-to-video	✅			MoE
Wan-AI	Wan2.1-VACE	video creation (T2V/I2V/FLF2V)	🙋			1.3B / 14B variants
Wan-AI	Wan2.2-S2V-14B	speech-to-video (image+audio→video)	✅	@xuechendi	#2751	720p, 81 frames; TP=2
Lightricks	LTX-2-T2V	text-to-video	✅	@fywc	#3294	512x768, 121 frames
Lightricks	LTX-2-I2V	image-to-video	✅	@fywc	#3294	same model as T2V
Lightricks	LTX-2.3	text-to-video + audio	✅	@oglok	#2893	22B, ~62 GiB peak; BWE 48kHz vocoder
Helios	Helios-Base	T2V / I2V / V2V	✅	@lengrongfu		stage 1 only; 8xH800 or 8xH200
Helios	Helios-Mid	video denoising	🙋	@lengrongfu		multi-stage pyramid; 8xH800 or 8xH200
Helios	Helios-Distilled	video (few-step)	🙋	@lengrongfu		DMD distillation; 8xH800 or 8xH200
SII-GAIR	MagiHuman	video+audio with lip sync	🙋			TP=4 for 80GB, DiT MoE+T5
XuGuo699	DreamID-Omni	I2V with identity preservation	🙋	@fywc		~72 GiB
Text-to-Speech / Audio
Qwen	Qwen3-TTS	TTS serving	✅	@chzhang2021	#3130	12Hz variants (CustomVoice, VoiceDesign, Base)
FunAudioLLM	CosyVoice3	TTS	⏳	@Moore-Z	#3486	2-stage: talker + flow-matching, 0.5B
FishAudio	Fish Speech S2 Pro	TTS / voice cloning	✅	@menjiantong	#3193	4B dual-AR, 44.1kHz
Mistral AI	Voxtral TTS	TTS with voice presets	🙋			4B, voice cloning
K2-FSA	OmniVoice	multilingual TTS (600+ lang)	🙋	@fray1024		zero-shot, Qwen3-0.6B backbone; 1xL20 48GB
OpenBMB	VoxCPM2	TTS (diffusion AR)	🙋	@wjinxu		2B, 48kHz, 30+ lang
Xiaomi (MiMo)	MiMo-Audio	audio gen / TTS / ASR / dialogue	🙋			multi-task audio 7B
Stability AI	Stable-Audio-Open	audio gen (music, SFX)	🙋	@Ronnie-Rui		diffusion-based
HKUST	AudioX	T2A / V2A / TV2A / T2M / V2M / TV2M	✅	@zhangj1an	#2077	diffusion maf-mmdit; ~97 GiB VRAM
Tencent	Covo-Audio-Chat	audio chat (audio in, text+audio out)	✅	@Dnoob	#2293	7B LLM + BigVGAN; ~20 GiB

How to contribute (step by step)

1. Claim a row — comment on this issue with the row + your hardware (e.g. Qwen3-TTS, 1xL40S 48GB). Maintainer will mark the row ⏳ and assign you.
2. Copy the template — cp recipes/_template.md recipes/<Vendor>/<Model>.md (template merged in #2646).
3. Fill the required sections — model + task, tested HW, env (CUDA/ROCm/CANN versions), launch command, verification command + expected output, important flags, known limitations, links back to docs/ and examples/.
4. Test end-to-end on the hardware you claimed — recipes that aren't personally validated should not be merged.
5. Open PR titled [Recipe] <Vendor>/<Model> (or [Recipe] <Vendor>/<Model> (<HW>) if adding a HW section to an existing recipe), link this issue, and paste the launch command + verification output in the PR body. Include a test plan describing what you verified (e.g. specific prompts, expected outputs, latency/throughput numbers) and the test results showing the actual output or logs.

The initial recipes/ template has been merged #2646. We now welcome community contributions for practical "run model X on hardware Y for task Z" recipes in vllm-omni.

Please use the merged template as the starting point and organize recipes by model vendor, for example:

recipes/
  Qwen/
    Qwen3-Omni.md
    Qwen3-TTS.md
    Qwen2.5-Omni.md
  Tencent-Hunyuan/
    HunyuanVideo.md
  GLM/
    GLM-Image.md
  MiMo/
    MiMo-Audio.md
  Wan-AI/
    Wan2.2-T2V.md
  ...

Each recipe should include:

• Model name and task
• Tested hardware configuration
• Required environment
• Launch command
• Verification command or expected output
• Important flags or stage configs
• Known limitations
• Links back to canonical docs/ and runnable examples/

High Priority Starter Recipes

• recipes/Qwen/Qwen3-TTS.md

  • Suggested scope: text-to-speech serving with Qwen3-TTS
  • Suggested hardware sections: CUDA GPU, NPU if available
  • Link to existing Qwen3-TTS docs/examples where possible

• recipes/Qwen/Qwen2.5-Omni.md

  • Suggested scope: omni-modal chat or speech interaction
  • Suggested hardware sections: CUDA GPU, NPU if available
  • Link to existing Qwen2.5-Omni docs/examples where possible

• recipes/Tencent-Hunyuan/HunyuanVideo.md

  • Suggested scope: text-to-video or image-to-video generation
  • Include tested VRAM requirements and performance notes where possible

• recipes/GLM/GLM-Image.md

  • Suggested scope: image generation with GLM-Image
  • Include recommended generation settings and validation output

• recipes/MiMo/MiMo-Audio.md

  • Suggested scope: audio generation or speech-related serving
  • Include model-specific setup and verification notes

Hardware Coverage We Want

• NVIDIA CUDA recipes

  • Example targets: A100 80GB, H100, L40S, RTX 4090 where applicable

• AMD ROCm recipes

  • Include ROCm version, GPU model, and any known caveats

• Intel XPU recipes

  • Suggested devices from community feedback:
  • Intel Arc Pro B50, 16GB
  • Intel Arc Pro B60, 24GB
  • Intel Arc Pro B70, 32GB

• Huawei NPU recipes

  • Include CANN/runtime versions and model-specific limitations

Good First Contributions

• Add a recipe for a model you have successfully run locally
• Add another hardware section to an existing recipe
• Add verification output to an existing recipe
• Link an existing recipe to the relevant docs/ and examples/
• Improve clarity around memory usage, flags, or known limitations

Contribution Guidelines

When opening a recipe PR, please include:

• The exact command you used
• Hardware details
• Software/runtime versions
• Whether the recipe was personally tested
• Test plan: what you verified (specific prompts, expected outputs, latency/throughput numbers)
• Test results: actual output or logs from your run
• Any limitations or assumptions
• Links to relevant examples or docs

Recipes should not replace canonical documentation. They should act as practical, community-maintained runbooks that point users to the right docs and runnable examples.

---

Motivation.

We'd like to propose a new community-maintained recipes/ area in vllm-project/vllm-omni to answer a recurring user question:

How do I run model X on hardware Y for task Z?

Today, users often struggle to find the best path for a concrete deployment scenario in vLLM-Omni.

There are a few reasons:

1. We currently do not have a dedicated place for operational runbooks that map a specific model, hardware target, and task to a known-good setup.
2. The current examples/ layout mixes model-specific and task-specific directories, which can be confusing for discovery.
3. We want to align the user experience with vllm-project/recipes, which already provides this style of practical guide for upstream vLLM.

Proposed Change.

Introduce a top-level recipes/ directory in vLLM-Omni for community-maintained runbooks.

To align with upstream vllm-project/recipes, recipes should be grouped by model vendor at the top level, with one Markdown file per model family by default.

Example direction:

recipes/
  Qwen/
    Qwen3-Omni.md
    Qwen3-TTS.md
    Qwen2.5-Omni.md
  Tencent-Hunyuan/
    HunyuanVideo.md
  GLM/
    GLM-Image.md
  MiMo/
    MiMo-Audio.md

Within each model doc, we should include multiple hardware-specific sections in the same Markdown file, following the structure used by the upstream DeepSeek guide: https://github.com/vllm-project/recipes/blob/main/DeepSeek/DeepSeek-V3.md

For example, a single recipe doc could contain sections such as:

• 1x A100 80GB
• 2x L40S
• 4x H100
• ROCm / NPU variants where applicable

Each section would describe:

• supported task(s)
• tested hardware
• required environment
• launch commands
• important flags / stage configs
• verification steps
• known limitations

Design Principles

• Align the top-level organization with vllm-project/recipes where practical, so discovery feels familiar across vLLM and vLLM-Omni.
• Keep one Markdown file per model family by default.
• Keep multiple hardware configurations inside the same recipe document unless the document becomes too large or hard to maintain.
• Use recipes for practical "known-good setup" guidance, not as the canonical source of product documentation.

Relationship to examples/

The current examples/ folder is still valuable, but it serves a different purpose:

• examples/: runnable code and scripts
• docs/: canonical documentation
• recipes/: practical "known-good setup" guides for concrete user scenarios

One benefit of adding recipes/ is that it may reduce pressure to make examples/ itself carry all discovery and onboarding needs.

This also gives us a cleaner answer to users who ask for a concrete deployment path without requiring them to infer it from a mix of task-oriented and model-oriented example folders.

Scope

This proposal is for community recipes only.

It is not intended to replace:

• canonical product documentation under docs/
• source-of-truth runnable examples under examples/

Instead, recipes would act as a user-oriented entry point that links back to those canonical materials.

Open Questions

1. Should recipe ownership be fully community-maintained, or should each recipe have one or two named maintainers?
2. Should recipes live only in the repo at first, or also be surfaced in the documentation site under a Community section?
3. Should we add a lightweight recipe template from the beginning to keep structure consistent?
4. Should any future cleanup of examples/ be handled in a separate RFC after recipes/ is established?

Initial Success Criteria

• A new user can quickly find a concrete recipe for a target model, hardware, and task.
• Recipes follow a consistent structure.
• Recipes are organized by vendor in a way that feels familiar to users of upstream vllm-project/recipes.
• Recipes link to examples/ and docs/ instead of duplicating canonical content.
• The approach improves user experience without adding confusion about where canonical documentation lives.

Suggested First Recipes

• recipes/Qwen/Qwen3-Omni.md
• recipes/Qwen/Qwen3-TTS.md
• recipes/Qwen/Qwen2.5-Omni.md
• recipes/Tencent-Hunyuan/HunyuanVideo.md
• recipes/GLM/GLM-Image.md
• recipes/MiMo/MiMo-Audio.md

Feedback welcome on structure, ownership, and how tightly we should align with upstream vllm-project/recipes.

Feedback Period.

CC List.

vllm-omni maintainer team @ywang96 @Gaohan123 @ZJY0516 @princepride @lishunyang12 .....

Any Other Things.