Nunchaku March Development Roadmap

[lmxyy]

Nunchaku March Development Roadmap

Hi everyone,

Thank you for trying out SVDQuant and Nunchaku! We know you may have encountered some issues, but rest assured—these are temporary. We're actively improving the codebase to make it more user-friendly and stable.

This roadmap outlines our key development goals for March 2025. Contributions and feedback are always welcome!

Focus Areas for March

• Enhance user experience and improve stability.
• Expand support for FLUX-based models and LoRAs.

Installation

• Windows (@lmxyy)

  • Build and distribute wheels
  • Provide official installers
  • Installation and usage guide video

• Linux/WSL (@lmxyy)

  • Build and distribute wheels
  • Develop a Docker image
  • Create an installation and usage guide video

LoRA

• Support high-rank LoRAs (@sxtyzhangzk)
• Enable composition of multiple LoRAs (@lmxyy)
• Support LoRA removal (@sxtyzhangzk, @lmxyy)
• Fix LoRA-related bugs: (lora好像不兼容，完全没起作用 #185 Can't convert fluxgym lora to SVDQuant format #194 lora 切换bug #197 加速非常明显，但是LORA 不起作用。 #198 @lmxyy)
• Resolved tempfile loading issue ('utf-8' codec can't decode byte 0xc1 in position 0: invalid start byte #120 lora加载转换节点的使用方法 #170 Lora converter block fails on windows due to "could not open file in read-only mode" error #176 lora error #192, included in v0.2.0)
• Fix FLUX.1-tools LoRAs (转flux1-canny-dev-lora报错 #182, @lmxyy)

Quantization

• Release FLUX.1-Tools FP4 models (@synxlin)
• Release 4-bit Shuttle-Jaguar (@lmxyy)
• Simplify deepcompressor backend (@synxlin)
• Develop a new ComfyUI node for quantization (@lmxyy)
• Improve fidelity of the 4-bit text encoder (@synxlin)

Speed

• Implement First-Block Cache (WaveSpeed-like) (Add dynamic Caching when batch_size = 1 for flux model #138, thanks @ita9naiwa, to appear in v0.2.0)
• Support low-bit attention (SageAttention-like) (@sxtyzhangzk)

Memory

• Optimize memory usage when loading T5 (@synxlin)
• Fix CPU offloading speed issue in ComfyUI (@lmxyy)
• Clean memory cache when deleting models (@lmxyy)

Workflows

• Add a Redux workflow example (@lmxyy)
• Fix FLUX.1-fill collapse issue (@lmxyy)

Other Fixes & Improvements

• Fix the resolution issues (Still have problem with resolution (orthodox and unorthordox resolution) #187 Assertion 'k <= MAXK' failed #188 Passing txt_ids 3d torch.Tensor #196, @sxtyzhangzk)
• Improve the HuggingFace and Modelscope model documents (@lmxyy)

Some future features in plan

• ControlNet support
• ACE-plus support
• IPAdapter integration
• Wan2.1 Support.
• 8-bit model support
• Operator modularization
• Auto tests and deployment CI.

Acknowledgments

We borrow the roadmap template from SGLang.

[nitinmukesh]

Looking forward to the next release.

few request

1. Some of us are using outside Comfy (diffusers), please do add examples (in examples, the existing examples are really helpful) for (Develop a dedicated node for LoRA conversion)
2. And if this can be prioritized please. Wan2.1 Support. :)

Thank you for all your hard work. 👍

[ycyy]

Make LoRA Great Again

[truenorth8]

@lmxyy
Wan2.1 video support would be really awesome.
Is there a rough timeline for it? e.g. end of April?

btw it's really insane how much faster things generate with Nunchaku,
thanks for the work you guys have put in