[vllm-omni]: Omni Quant Support

[thuang6]

Feature Description

This epic is created to track Omni model (vllm-omni targeted) quantization support status.
main dtypes includes: INT4, MXFP8, MXFP4 is future target

Motivation and Use Case

omni model enable status:

model	architecture	owner	INT4	MXFP8
Qwen3-omni	Qwen3OmniMoeForConditionalGeneration	Liang	ok
Qwen2.5-omni	Qwen2_5OmniForConditionalGeneration	Liang	ok
HunyuanImage-3.0	HunyuanImage3ForCausalMM	Chang	ok
Z-Image	ZImageTransformer2DModel	Xin	ok
Next-Step	ZImageTransformer2DModel	Xin	WIP (Feature request in vllm-omni)
MiMo-Audio-7B-Instruct	MiMoAudioForCausalLM	Weiwei	fail
Qwen3-TTS-12Hz-1.7B-CustomVoice	Qwen3TTSForConditionalGeneration	Weiwei	fail
GLM-Image	GlmImageForConditionalGeneration GlmImageTransformer2DModel	Liang	ok
BAGEL-7B-MoT	BagelForConditionalGeneration	Liang
Ovis-Image	OvisImageTransformer2DModel	Liang

Alternatives Considered

No response

Definition of Done

No response

Additional Context

No response

[mengniwang95]

#1519 diffusion model saving is done

[changwangss]

#1502 merged, HunyuanImage-3.0 is ready for auto-round.