You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
**Nunchaku** is a high-performance inference engine optimized for 4-bit neural networks, as introduced in our paper [SVDQuant](http://arxiv.org/abs/2411.05007). For the underlying quantization library, check out [DeepCompressor](https://github.com/mit-han-lab/deepcompressor).
12
+
**Nunchaku** is a high-performance inference engine optimized for 4-bit neural networks, as introduced in our paper [SVDQuant](http://arxiv.org/abs/2411.05007). For the underlying quantization library, check out [DeepCompressor](https://github.com/nunchaku-tech/deepcompressor).
13
13
14
-
Join our user groups on [**Slack**](https://join.slack.com/t/nunchaku/shared_invite/zt-3170agzoz-NgZzWaTrEj~n2KEV3Hpl5Q), [**Discord**](https://discord.gg/Wk6PnwX9Sm) and [**WeChat**](https://huggingface.co/datasets/nunchaku-tech/cdn/resolve/main/nunchaku/assets/wechat.jpg) to engage in discussions with the community! More details can be found [here](https://github.com/mit-han-lab/nunchaku/issues/149). If you have any questions, run into issues, or are interested in contributing, don’t hesitate to reach out!
14
+
Join our user groups on [**Slack**](https://join.slack.com/t/nunchaku/shared_invite/zt-3170agzoz-NgZzWaTrEj~n2KEV3Hpl5Q), [**Discord**](https://discord.gg/Wk6PnwX9Sm) and [**WeChat**](https://huggingface.co/datasets/nunchaku-tech/cdn/resolve/main/nunchaku/assets/wechat.jpg) to engage in discussions with the community! More details can be found [here](https://github.com/nunchaku-tech/nunchaku/issues/149). If you have any questions, run into issues, or are interested in contributing, don’t hesitate to reach out!
15
15
16
16
## News
17
17
18
18
-**[2025-07-13]** 🚀 The official [**Nunchaku documentation**](https://nunchaku.tech/docs/nunchaku/) is now live! Explore comprehensive guides and resources to help you get started.
19
19
-**[2025-06-29]** 🔥 Support **FLUX.1-Kontext**! Try out our [example script](./examples/flux.1-kontext-dev.py) to see it in action! Our demo is available at this [link](https://svdquant.mit.edu/kontext/)!
20
-
-**[2025-06-01]** 🚀 **Release v0.3.0!** This update adds support for multiple-batch inference, [**ControlNet-Union-Pro 2.0**](https://huggingface.co/Shakker-Labs/FLUX.1-dev-ControlNet-Union-Pro-2.0), initial integration of [**PuLID**](https://github.com/ToTheBeginning/PuLID), and introduces [**Double FB Cache**](examples/flux.1-dev-double_cache.py). You can now load Nunchaku FLUX models as a single file, and our upgraded [**4-bit T5 encoder**](https://huggingface.co/mit-han-lab/nunchaku-t5) now matches **FP8 T5** in quality!
20
+
-**[2025-06-01]** 🚀 **Release v0.3.0!** This update adds support for multiple-batch inference, [**ControlNet-Union-Pro 2.0**](https://huggingface.co/Shakker-Labs/FLUX.1-dev-ControlNet-Union-Pro-2.0), initial integration of [**PuLID**](https://github.com/ToTheBeginning/PuLID), and introduces [**Double FB Cache**](examples/flux.1-dev-double_cache.py). You can now load Nunchaku FLUX models as a single file, and our upgraded [**4-bit T5 encoder**](https://huggingface.co/nunchaku-tech/nunchaku-t5) now matches **FP8 T5** in quality!
21
21
-**[2025-04-16]** 🎥 Released tutorial videos in both [**English**](https://youtu.be/YHAVe-oM7U8?si=cM9zaby_aEHiFXk0) and [**Chinese**](https://www.bilibili.com/video/BV1BTocYjEk5/?share_source=copy_web&vd_source=8926212fef622f25cc95380515ac74ee) to assist installation and usage.
22
-
-**[2025-04-09]** 📢 Published the [April roadmap](https://github.com/mit-han-lab/nunchaku/issues/266) and an [FAQ](https://github.com/mit-han-lab/nunchaku/discussions/262) to help the community get started and stay up to date with Nunchaku’s development.
22
+
-**[2025-04-09]** 📢 Published the [April roadmap](https://github.com/nunchaku-tech/nunchaku/issues/266) and an [FAQ](https://github.com/nunchaku-tech/nunchaku/discussions/262) to help the community get started and stay up to date with Nunchaku’s development.
23
23
-**[2025-04-05]** 🚀 **Nunchaku v0.2.0 released!** This release brings [**multi-LoRA**](examples/flux.1-dev-multiple-lora.py) and [**ControlNet**](examples/flux.1-dev-controlnet-union-pro.py) support with even faster performance powered by [**FP16 attention**](#fp16-attention) and [**First-Block Cache**](#first-block-cache). We've also added compatibility for [**20-series GPUs**](examples/flux.1-dev-turing.py) — Nunchaku is now more accessible than ever!
24
24
25
25
<details>
@@ -32,14 +32,14 @@ Join our user groups on [**Slack**](https://join.slack.com/t/nunchaku/shared_inv
32
32
-**[2025-02-04]****🚀 4-bit [FLUX.1-tools](https://blackforestlabs.ai/flux-1-tools/) is here!** Enjoy a **2-3× speedup** over the original models. Check out the [examples](./examples) for usage. **ComfyUI integration is coming soon!**
33
33
-**[2025-01-23]** 🚀 **4-bit [SANA](https://nvlabs.github.io/Sana/) support is here!** Experience a 2-3× speedup compared to the 16-bit model. Check out the [usage example](examples/sana1.6b_pag.py) and the [deployment guide](app/sana/t2i) for more details. Explore our live demo at [svdquant.mit.edu](https://svdquant.mit.edu)!
34
34
-**[2025-01-22]** 🎉 [**SVDQuant**](http://arxiv.org/abs/2411.05007) has been accepted to **ICLR 2025**!
35
-
-**[2024-12-08]** Support [ComfyUI](https://github.com/comfyanonymous/ComfyUI). Please check [mit-han-lab/ComfyUI-nunchaku](https://github.com/mit-han-lab/ComfyUI-nunchaku) for the usage.
36
-
-**[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.
35
+
-**[2024-12-08]** Support [ComfyUI](https://github.com/comfyanonymous/ComfyUI). Please check [ComfyUI-nunchaku](https://github.com/nunchaku-tech/ComfyUI-nunchaku) for the usage.
36
+
-**[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/nunchaku-tech/deepcompressor) for the quantization library.
**Nunchaku** is a high-performance inference engine for low-bit neural networks. It implements **SVDQuant**, 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.
44
44
45
45
**SVDQuant: Absorbing Outliers by Low-Rank Components for 4-Bit Diffusion Models**<br>
Overview of SVDQuant. Stage1: Originally, both the activation $\boldsymbol{X}$ and weights $\boldsymbol{W}$ contain outliers, making 4-bit quantization challenging. Stage 2: We migrate the outliers from activations to weights, resulting in the updated activation $\hat{\boldsymbol{X}}$ and weights $\hat{\boldsymbol{W}}$. While $\hat{\boldsymbol{X}}$ becomes easier to quantize, $\hat{\boldsymbol{W}}$ now becomes more difficult. Stage 3: SVDQuant further decomposes $\hat{\boldsymbol{W}}$ into a low-rank component $\boldsymbol{L}_1\boldsymbol{L}_2$ and a residual $\hat{\boldsymbol{W}}-\boldsymbol{L}_1\boldsymbol{L}_2$ with SVD. Thus, the quantization difficulty is alleviated by the low-rank branch, which runs at 16-bit precision.
55
+
Overview of SVDQuant. Stage1: Originally, both the activation $\boldsymbol{X}$ and weights $\boldsymbol{W}$ contain outliers, making 4-bit quantization challenging. Stage 2: We migrate the outliers from activations to weights, resulting in the updated activation $\hat{\boldsymbol{X}}$ and weights $\hat{\boldsymbol{W}}$. While $\hat{\boldsymbol{X}}$ becomes easier to quantize, $\hat{\boldsymbol{W}}$ now becomes more difficult. Stage 3: SVDQuant further decomposes $\hat{\boldsymbol{W}}$ into a low-rank component $\boldsymbol{L}_1\boldsymbol{L}_2$ and a residual $\hat{\boldsymbol{W}}-\boldsymbol{L}_1\boldsymbol{L}_2$ with SVD. Thus, the quantization difficulty is alleviated by the low-rank branch, which runs at 16-bit precision.
56
56
57
57
#### Nunchaku Engine Design
58
58
59
-
 (a) Naïvely running low-rank branch with rank 32 will introduce 57% latency overhead due to extra read of 16-bit inputs in *Down Projection* and extra write of 16-bit outputs in *Up Projection*. Nunchaku optimizes this overhead with kernel fusion. (b) *Down Projection* and *Quantize* kernels use the same input, while *Up Projection* and *4-Bit Compute* kernels share the same output. To reduce data movement overhead, we fuse the first two and the latter two kernels together.
59
+
 (a) Naïvely running low-rank branch with rank 32 will introduce 57% latency overhead due to extra read of 16-bit inputs in *Down Projection* and extra write of 16-bit outputs in *Up Projection*. Nunchaku optimizes this overhead with kernel fusion. (b) *Down Projection* and *Quantize* kernels use the same input, while *Up Projection* and *4-Bit Compute* kernels share the same output. To reduce data movement overhead, we fuse the first two and the latter two kernels together.
60
60
61
61
## Performance
62
62
63
-
SVDQuant reduces the 12B FLUX.1 model size by 3.6× and cuts the 16-bit model's memory usage by 3.5×. With Nunchaku, our INT4 model runs 3.0× faster than the NF4 W4A16 baseline on both desktop and laptop NVIDIA RTX 4090 GPUs. Notably, on the laptop 4090, it achieves a total 10.1× speedup by eliminating CPU offloading. Our NVFP4 model is also 3.1× faster than both BF16 and NF4 on the RTX 5090 GPU.
63
+
SVDQuant reduces the 12B FLUX.1 model size by 3.6× and cuts the 16-bit model's memory usage by 3.5×. With Nunchaku, our INT4 model runs 3.0× faster than the NF4 W4A16 baseline on both desktop and laptop NVIDIA RTX 4090 GPUs. Notably, on the laptop 4090, it achieves a total 10.1× speedup by eliminating CPU offloading. Our NVFP4 model is also 3.1× faster than both BF16 and NF4 on the RTX 5090 GPU.
Please check [here](https://github.com/mit-han-lab/nunchaku/issues/431) for the roadmap for the Summer.
79
+
Please check [here](https://github.com/nunchaku-tech/nunchaku/issues/431) for the roadmap for the Summer.
79
80
80
81
## Contact Us
81
82
@@ -116,4 +117,4 @@ Nunchaku is also inspired by many open-source libraries, including (but not limi
116
117
117
118
## Star History
118
119
119
-
[](https://www.star-history.com/#mit-han-lab/nunchaku&Date)
120
+
[](https://www.star-history.com/#nunchaku-tech/nunchaku&Date)
0 commit comments