[Feature] Z-Image-Turbo Support

[atgctg]

Checklist

• 1. If the issue you raised is not a feature but a question, please raise a discussion at https://github.com/mit-han-lab/nunchaku/discussions/new/choose. Otherwise, it will be closed.
• 2. I will do my best to describe the issue in English.

Motivation

Support a quantized FP4 version of the popular Tongyi-MAI/Z-Image-Turbo image model in Nunchaku.

Related resources

No response

[tazztone]

ig z-image is architecturally similar to qwen-image. so we might be in luck

[vuaieo]

that would be very cool

[alex-mitov]

Do we really need a quantized version of a model that already runs on a potato PCs (by today's standards)? Seems like needless waste of resources. It takes several seconds on mid-range computers to generate an image. Flux.2 is much more logical candidate for quantization given its heavy size and lengthy generation times of a couple of minutes on upper mid-range PCs, and it's basically unusable on mid-range hardware and below.

[juntaosun]

They seem to have reduced their releases and updates; the last time was three weeks ago.

[frankyifei]

It is still valuable to have a quantized version when you need to use lora or wanting to generate large resolution. Z-image handles 2k output very well but it takes more vram for that. Control net and other essential tools also requires large vram not every one has

[liuguicen]

+1

[Milor123]

Yeahh +1

[IceSandwich]

Still need a quantized model to reduce the inference cost even it has already run in "a potato pc". Would like to see its performance in realtime generation.

[bukit-kronik]

Yeah real-time generation is the keyword here.

[hemantp66]

+1

[vgabbo]

+1

[Yasand123]

Do we really need a quantized version of a model that already runs on a potato PCs (by today's standards)? Seems like needless waste of resources. It takes several seconds on mid-range computers to generate an image. Flux.2 is much more logical candidate for quantization given its heavy size and lengthy generation times of a couple of minutes on upper mid-range PCs, and it's basically unusable on mid-range hardware and below.

2060 super (8gb) generates a 1024x1024 image in 120 seconds. My GPU isn't the best sure but it's not the worst either, it's still RTX. I definitively think nunchaku ZIT would be a big help.

[badraymen]

+1

[fappaz]

The quantized models have been uploaded to hugging face, so we should expect it to have the proper support very soon.