Nunchaku is an inference engine designed for 4-bit diffusion models, as demonstrated in our paper [SVDQuant](http://arxiv.org/abs/2411.05007). Please check [DeepCompressor](https://github.com/mit-han-lab/deepcompressor) for the quantization library.
**Nunchaku** is an efficient inference engine designed for 4-bit diffusion models, as demonstrated in our paper [SVDQuant](http://arxiv.org/abs/2411.05007). Please check [DeepCompressor](https://github.com/mit-han-lab/deepcompressor) for the quantization library.
Check [here](https://github.com/mit-han-lab/nunchaku/issues/149) to join our user groups on [**Slack**](https://join.slack.com/t/nunchaku/shared_invite/zt-3170agzoz-NgZzWaTrEj~n2KEV3Hpl5Q) and [**WeChat**](./assets/wechat.jpg) for discussions! If you have any questions, encounter issues, or are interested in contributing to the codebase, feel free to share your thoughts there!
-**[2025-03-11]****🚀 Release [4-bit Shuttle-Jaguar](https://huggingface.co/mit-han-lab/svdq-int4-shuttle-jaguar)!** Check the INT4 models in our [HuggingFace](https://huggingface.co/collections/mit-han-lab/svdquant-67493c2c2e62a1fc6e93f45c) or [ModelScope](https://modelscope.cn/collections/svdquant-468e8f780c2641) collections! FP4 models are coming soon!
-**[2025-03-07]** 🚀 **Nunchaku v0.1.4 Released!** We've supported [4-bit text encoder and per-layer CPU offloading](#Low-Memory-Inference), reducing FLUX's minimum memory requirement to just **4 GiB** while maintaining a **2–3× speedup**. This update also fixes various issues related to resolution, LoRA, pin memory, and runtime stability. Check out the release notes for full details!
...
...
@@ -19,6 +24,8 @@ Check [here](https://github.com/mit-han-lab/nunchaku/issues/149) to join our use
-**[2024-12-08]** Support [ComfyUI](https://github.com/comfyanonymous/ComfyUI). Please check [comfyui/README.md](comfyui/README.md) for the usage.
-**[2024-11-07]** 🔥 Our latest **W4A4** Diffusion model quantization work [**SVDQuant**](https://hanlab.mit.edu/projects/svdquant) is publicly released! Check [**DeepCompressor**](https://github.com/mit-han-lab/deepcompressor) for the quantization library.
## Overview
SVDQuant is a post-training quantization technique for 4-bit weights and activations that well maintains visual fidelity. On 12B FLUX.1-dev, it achieves 3.6× memory reduction compared to the BF16 model. By eliminating CPU offloading, it offers 8.7× speedup over the 16-bit model when on a 16GB laptop 4090 GPU, 3× faster than the NF4 W4A16 baseline. On PixArt-∑, it demonstrates significantly superior visual quality over other W4A4 or even W4A8 baselines. "E2E" means the end-to-end latency including the text encoder and VAE decoder.
