README.md

<div align="center" id="nunchaku_logo">
  <img src="assets/nunchaku.svg" alt="logo" width="220"></img>
</div>
<h3 align="center">
<a href="http://arxiv.org/abs/2411.05007"><b>Paper</b></a> | <a href="https://hanlab.mit.edu/projects/svdquant"><b>Website</b></a> | <a href="https://hanlab.mit.edu/blog/svdquant"><b>Blog</b></a> | <a href="https://svdquant.mit.edu"><b>Demo</b></a> | <a href="https://huggingface.co/collections/mit-han-lab/svdquant-67493c2c2e62a1fc6e93f45c"><b>HuggingFace</b></a> | <a href="https://modelscope.cn/collections/svdquant-468e8f780c2641"><b>ModelScope</b></a> | <a href="https://github.com/mit-han-lab/ComfyUI-nunchaku"><b>ComfyUI</b></a>
</h3>


**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).

Join our user groups on [**Slack**](https://join.slack.com/t/nunchaku/shared_invite/zt-3170agzoz-NgZzWaTrEj~n2KEV3Hpl5Q) and [**WeChat**](./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!

## News

- **[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!
- **[2025-03-17]** 🚀 Released NVFP4 4-bit [Shuttle-Jaguar](https://huggingface.co/mit-han-lab/svdq-int4-shuttle-jaguar) and FLUX.1-tools and also upgraded the INT4 FLUX.1-tool models. Download and update your models from our [HuggingFace](https://huggingface.co/collections/mit-han-lab/svdquant-67493c2c2e62a1fc6e93f45c) or [ModelScope](https://modelscope.cn/collections/svdquant-468e8f780c2641) collections!
- **[2025-03-13]** 📦 Separate the ComfyUI node into a [standalone repository](https://github.com/mit-han-lab/ComfyUI-nunchaku) for easier installation and release node v0.1.6! Plus, [4-bit Shuttle-Jaguar](https://huggingface.co/mit-han-lab/svdq-int4-shuttle-jaguar) is now fully supported!
- **[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!
- **[2025-02-20]** 🚀 We release the [pre-built wheels](https://huggingface.co/mit-han-lab/nunchaku) to simplify installation! Check [here](#Installation) for the guidance!
- **[2025-02-20]** 🚀 **Support NVFP4 precision on NVIDIA RTX 5090!** NVFP4 delivers superior image quality compared to INT4, offering **~3× speedup** on the RTX 5090 over BF16. Learn more in our [blog](https://hanlab.mit.edu/blog/svdquant-nvfp4), checkout  [`examples`](./examples) for usage and try [our demo](https://svdquant.mit.edu/flux1-schnell/) online!
- **[2025-02-18]** 🔥 [**Customized LoRA conversion**](#Customized-LoRA) and [**model quantization**](#Customized-Model-Quantization) instructions are now available! **[ComfyUI](./comfyui)** workflows now support **customized LoRA**, along with **FLUX.1-Tools**!
- **[2025-02-11]** 🎉 **[SVDQuant](http://arxiv.org/abs/2411.05007) has been selected as a ICLR 2025 Spotlight! FLUX.1-tools Gradio demos are now available!** Check [here](#gradio-demos) for the usage details! Our new [depth-to-image demo](https://svdquant.mit.edu/flux1-depth-dev/) is also online—try it out!


<details>
<summary>More</summary>

- **[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!**
- **[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/sana_1600m_pag.py) and the [deployment guide](app/sana/t2i) for more details. Explore our live demo at [svdquant.mit.edu](https://svdquant.mit.edu)!
- **[2025-01-22]** 🎉 [**SVDQuant**](http://arxiv.org/abs/2411.05007) has been accepted to **ICLR 2025**!
- **[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.
- **[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.

</details>

## Overview

![teaser](./assets/teaser.jpg)
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.

**SVDQuant: Absorbing Outliers by Low-Rank Components for 4-Bit Diffusion Models**<br>
[Muyang Li](https://lmxyy.me)\*, [Yujun Lin](https://yujunlin.com)\*, [Zhekai Zhang](https://hanlab.mit.edu/team/zhekai-zhang)\*, [Tianle Cai](https://www.tianle.website/#/), [Xiuyu Li](https://xiuyuli.com), [Junxian Guo](https://github.com/JerryGJX), [Enze Xie](https://xieenze.github.io), [Chenlin Meng](https://cs.stanford.edu/~chenlin/), [Jun-Yan Zhu](https://www.cs.cmu.edu/~junyanz/), and [Song Han](https://hanlab.mit.edu/songhan) <br>
*MIT, NVIDIA, CMU, Princeton, UC Berkeley, SJTU, and Pika Labs* <br>

<p align="center">
  <img src="assets/demo.gif" width="100%"/>
</p>

## Method

#### Quantization Method -- SVDQuant

![intuition](./assets/intuition.gif)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. 

#### Nunchaku Engine Design

![engine](./assets/engine.jpg) (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.


## Performance

![efficiency](./assets/efficiency.jpg)SVDQuant reduces the model size of the 12B FLUX.1 by 3.6×. Additionally, Nunchaku, further cuts memory usage of the 16-bit model by 3.5× and delivers 3.0× speedups over the NF4 W4A16 baseline on both the desktop and laptop NVIDIA RTX 4090 GPUs. Remarkably, on laptop 4090, it achieves in total 10.1× speedup by eliminating CPU offloading.

## Installation

### Wheels

#### Prerequisites
Before installation, ensure you have [PyTorch>=2.5](https://pytorch.org/) installed. For example, you can use the following command to install PyTorch 2.6:

```shell
pip install torch==2.6 torchvision==0.21 torchaudio==2.6
```

#### Install nunchaku
Once PyTorch is installed, you can directly install `nunchaku` from our wheel repositories on [Hugging Face](https://huggingface.co/mit-han-lab/nunchaku/tree/main) or [ModelScope](https://modelscope.cn/models/Lmxyy1999/nunchaku) or [GitHub release](https://github.com/mit-han-lab/nunchaku/releases). Be sure to select the appropriate wheel for your Python and PyTorch version. For example, for Python 3.11 and PyTorch 2.6:

```shell
pip install https://huggingface.co/mit-han-lab/nunchaku/resolve/main/nunchaku-0.2.0+torch2.6-cp311-cp311-linux_x86_64.whl
```

**Note**: If you're using a Blackwell GPU (e.g., 50-series GPUs), install a wheel with PyTorch 2.7. Additionally, use **FP4 models** instead of INT4 models.

### Build from Source

**Note**:

*  Make sure your CUDA version is **at least 12.2 on Linux** and **at least 12.6 on Windows**. If you're using a Blackwell GPU (e.g., 50-series GPUs), CUDA **12.8 or higher is required**.

*  For Windows users, please refer to [this issue](https://github.com/mit-han-lab/nunchaku/issues/6) for the instruction. Please upgrade your MSVC compiler to the latest version.

*  We currently support only NVIDIA GPUs with architectures sm_75 (Turing: RTX 2080), sm_86 (Ampere: RTX 3090, A6000), sm_89 (Ada: RTX 4090), and sm_80 (A100). See [this issue](https://github.com/mit-han-lab/nunchaku/issues/1) for more details.


1. Install dependencies:

   ```shell
   conda create -n nunchaku python=3.11
   conda activate nunchaku
   pip install torch torchvision torchaudio
   pip install ninja wheel diffusers transformers accelerate sentencepiece protobuf huggingface_hub
   
   # For gradio demos
   pip install peft opencv-python gradio spaces GPUtil  
   ```

   To enable NVFP4 on Blackwell GPUs (e.g., 50-series GPUs), please install nightly PyTorch with CUDA 12.8. The installation command can be:

   ```shell
   pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128
   ```

2. Install `nunchaku` package:
    Make sure you have `gcc/g++>=11`. If you don't, you can install it via Conda on Linux:

    ```shell
    conda install -c conda-forge gxx=11 gcc=11
    ```

    For Windows users, you can download and install the lastest [Visual Studio](https://visualstudio.microsoft.com/thank-you-downloading-visual-studio/?sku=Community&channel=Release&version=VS2022&source=VSLandingPage&cid=2030&passive=false).
    
    Then build the package from source with
    
    ```shell
    git clone https://github.com/mit-han-lab/nunchaku.git
    cd nunchaku
    git submodule init
    git submodule update
    python setup.py develop
    ```
    
    If you are building wheels for distribution, use:
    
    ```shell
    NUNCHAKU_INSTALL_MODE=ALL NUNCHAKU_BUILD_WHEELS=1 python -m build --wheel --no-isolation
    ```
    
    Make sure to set the environment variable `NUNCHAKU_INSTALL_MODE` to `ALL`. Otherwise, the generated wheels will only work on GPUs with the same architecture as the build machine.

### Docker (Coming soon)

**[Optional]** You can verify your installation by running: `python -m nunchaku.test`. This command will download and run our 4-bit FLUX.1-schnell model.

## Usage Example

In [examples](examples), we provide minimal scripts for running INT4 [FLUX.1](https://github.com/black-forest-labs/flux) and [SANA](https://github.com/NVlabs/Sana) models with Nunchaku. It shares the same APIs as [diffusers](https://github.com/huggingface/diffusers) and can be used in a similar way. For example, the [script](examples/flux.1-dev.py) for [FLUX.1-dev](https://huggingface.co/black-forest-labs/FLUX.1-dev) is as follows:

```python
import torch
from diffusers import FluxPipeline

from nunchaku import NunchakuFluxTransformer2dModel
from nunchaku.utils import get_precision

precision = get_precision()  # auto-detect your precision is 'int4' or 'fp4' based on your GPU
transformer = NunchakuFluxTransformer2dModel.from_pretrained(f"mit-han-lab/svdq-{precision}-flux.1-dev")
pipeline = FluxPipeline.from_pretrained(
    "black-forest-labs/FLUX.1-dev", transformer=transformer, torch_dtype=torch.bfloat16
).to("cuda")
image = pipeline("A cat holding a sign that says hello world", num_inference_steps=50, guidance_scale=3.5).images[0]
image.save(f"flux.1-dev-{precision}.png")
```

**Note**: If you're using a **Turing GPU (e.g., NVIDIA 20-series)**, make sure to set `torch_dtype=torch.float16` and use our `nunchaku-fp16` attention module as below. A complete example is available in [`examples/flux.1-dev-turing.py`](examples/flux.1-dev-turing.py).

### FP16 Attention

In addition to FlashAttention-2, Nunchaku introduces a custom FP16 attention implementation that achieves up to **1.2× faster performance** on NVIDIA 30-, 40-, and even 50-series GPUs—without loss in precision. To enable it, simply use:

```python
transformer.set_attention_impl("nunchaku-fp16")
```

See [`examples/flux.1-dev-fp16attn.py`](examples/flux.1-dev-fp16attn.py) for a complete example.

### First-Block Cache

Nunchaku supports [First-Block Cache](https://github.com/chengzeyi/ParaAttention?tab=readme-ov-file#first-block-cache-our-dynamic-caching) to accelerate long-step denoising. Enable it easily with:

```python
apply_cache_on_pipe(pipeline, residual_diff_threshold=0.12)
```

You can tune the `residual_diff_threshold` to balance speed and quality: larger values yield faster inference at the cost of some quality. A recommended value is `0.12`, which provides up to **2× speedup** for 50-step denoising and **1.4× speedup** for 30-step denoising. See the full example in [`examples/flux.1-dev-cache.py`](examples/flux.1-dev-cache.py).

### CPU Offloading

To minimize GPU memory usage, Nunchaku supports CPU offloading—requiring as little as **4 GiB** of GPU memory. You can enable it by setting `offload=True` when initializing `NunchakuFluxTransformer2dModel`, and then calling:

```python
pipeline.enable_sequential_cpu_offload()
```

For a complete example, refer to [`examples/flux.1-dev-offload.py`](examples/flux.1-dev-offload.py).

## Customized LoRA

![lora](./assets/lora.jpg)

[SVDQuant](http://arxiv.org/abs/2411.05007) seamlessly integrates with off-the-shelf LoRAs without requiring requantization. You can simply use your LoRA with:

```python
transformer.update_lora_params(path_to_your_lora)
transformer.set_lora_strength(lora_strength)
```

`path_to_your_lora` can also be a remote HuggingFace path. In [`examples/flux.1-dev-lora.py`](examples/flux.1-dev-lora.py), we provide a minimal example script for running [Ghibsky](https://huggingface.co/aleksa-codes/flux-ghibsky-illustration) LoRA with SVDQuant's 4-bit FLUX.1-dev:

```python
import torch
from diffusers import FluxPipeline

from nunchaku import NunchakuFluxTransformer2dModel
from nunchaku.utils import get_precision

precision = get_precision()  # auto-detect your precision is 'int4' or 'fp4' based on your GPU
transformer = NunchakuFluxTransformer2dModel.from_pretrained(f"mit-han-lab/svdq-{precision}-flux.1-dev")
pipeline = FluxPipeline.from_pretrained(
    "black-forest-labs/FLUX.1-dev", transformer=transformer, torch_dtype=torch.bfloat16
).to("cuda")

### LoRA Related Code ###
transformer.update_lora_params(
    "aleksa-codes/flux-ghibsky-illustration/lora.safetensors"
)  # Path to your LoRA safetensors, can also be a remote HuggingFace path
transformer.set_lora_strength(1)  # Your LoRA strength here
### End of LoRA Related Code ###

image = pipeline(
    "GHIBSKY style, cozy mountain cabin covered in snow, with smoke curling from the chimney and a warm, inviting light spilling through the windows",  # noqa: E501
    num_inference_steps=25,
    guidance_scale=3.5,
).images[0]
image.save(f"flux.1-dev-ghibsky-{precision}.png")
```

To compose multiple LoRAs, you can use `nunchaku.lora.flux.compose.compose_lora` to compose them. The usage is 

```python
composed_lora = compose_lora(
    [
        ("PATH_OR_STATE_DICT_OF_LORA1", lora_strength1),
        ("PATH_OR_STATE_DICT_OF_LORA2", lora_strength2),
        # Add more LoRAs as needed
    ]
)  # set your lora strengths here when using composed lora
transformer.update_lora_params(composed_lora)
```

You can specify individual strengths for each LoRA in the list. For a complete example, refer to [`examples/flux.1-dev-multiple-lora.py`](examples/flux.1-dev-multiple-lora.py).

**For ComfyUI users, you can directly use our LoRA loader. The converted LoRA is deprecated. Please refer to [mit-han-lab/ComfyUI-nunchaku](https://github.com/mit-han-lab/ComfyUI-nunchaku) for more details.**

## ControlNets

Nunchaku supports both the [FLUX.1-tools](https://blackforestlabs.ai/flux-1-tools/) and the [FLUX.1-dev-ControlNet-Union-Pro](https://huggingface.co/Shakker-Labs/FLUX.1-dev-ControlNet-Union-Pro) models. Example scripts can be found in the [`examples`](examples) directory.

![control](./assets/control.jpg)

## ComfyUI

Please refer to [mit-han-lab/ComfyUI-nunchaku](https://github.com/mit-han-lab/ComfyUI-nunchaku) for the usage in [ComfyUI](https://github.com/comfyanonymous/ComfyUI).

## Gradio Demos

* FLUX.1 Models
  * Text-to-image: see [`app/flux.1/t2i`](app/flux.1/t2i).
  * Sketch-to-Image ([pix2pix-Turbo](https://github.com/GaParmar/img2img-turbo)): see [`app/flux.1/sketch`](app/flux.1/sketch).
  * Depth/Canny-to-Image ([FLUX.1-tools](https://blackforestlabs.ai/flux-1-tools/)): see [`app/flux.1/depth_canny`](app/flux.1/depth_canny).
  * Inpainting ([FLUX.1-Fill-dev](https://huggingface.co/black-forest-labs/FLUX.1-Depth-dev)): see [`app/flux.1/fill`](app/flux.1/fill).
  * Redux ([FLUX.1-Redux-dev](https://huggingface.co/black-forest-labs/FLUX.1-Redux-dev)): see [`app/flux.1/redux`](app/flux.1/redux).
* SANA:
  * Text-to-image: see [`app/sana/t2i`](app/sana/t2i).

## Customized Model Quantization

Please refer to [mit-han-lab/deepcompressor](https://github.com/mit-han-lab/deepcompressor/tree/main/examples/diffusion). A simpler workflow is coming soon.

## Benchmark

Please refer to [app/flux/t2i/README.md](app/flux/t2i/README.md) for instructions on reproducing our paper's quality results and benchmarking inference latency on FLUX.1 models.

## Roadmap

Please check [here](https://github.com/mit-han-lab/nunchaku/issues/201) for the roadmap for March.

## Citation

If you find `nunchaku` useful or relevant to your research, please cite our paper:

```bibtex
@inproceedings{
  li2024svdquant,
  title={SVDQuant: Absorbing Outliers by Low-Rank Components for 4-Bit Diffusion Models},
  author={Li*, Muyang and Lin*, Yujun and Zhang*, Zhekai and Cai, Tianle and Li, Xiuyu and Guo, Junxian and Xie, Enze and Meng, Chenlin and Zhu, Jun-Yan and Han, Song},
  booktitle={The Thirteenth International Conference on Learning Representations},
  year={2025}
}
```

## Related Projects

* [Efficient Spatially Sparse Inference for Conditional GANs and Diffusion Models](https://arxiv.org/abs/2211.02048), NeurIPS 2022 & T-PAMI 2023
* [SmoothQuant: Accurate and Efficient Post-Training Quantization for Large Language Models](https://arxiv.org/abs/2211.10438), ICML 2023
* [Q-Diffusion: Quantizing Diffusion Models](https://arxiv.org/abs/2302.04304), ICCV 2023
* [AWQ: Activation-aware Weight Quantization for LLM Compression and Acceleration](https://arxiv.org/abs/2306.00978), MLSys 2024
* [DistriFusion: Distributed Parallel Inference for High-Resolution Diffusion Models](https://arxiv.org/abs/2402.19481), CVPR 2024
* [QServe: W4A8KV4 Quantization and System Co-design for Efficient LLM Serving](https://arxiv.org/abs/2405.04532), MLSys 2025
* [SANA: Efficient High-Resolution Image Synthesis with Linear Diffusion Transformers](https://arxiv.org/abs/2410.10629), ICLR 2025

## Contact Us
For enterprises interested in adopting SVDQuant or Nunchaku, including technical consulting, sponsorship opportunities, or partnership inquiries, please contact us at muyangli@mit.edu.

## Acknowledgments

We thank MIT-IBM Watson AI Lab, MIT and Amazon Science Hub, MIT AI Hardware Program, National Science Foundation, Packard Foundation, Dell, LG, Hyundai, and Samsung for supporting this research. We thank NVIDIA for donating the DGX server.

We use [img2img-turbo](https://github.com/GaParmar/img2img-turbo) to train the sketch-to-image LoRA. Our text-to-image and sketch-to-image UI is built upon [playground-v.25](https://huggingface.co/spaces/playgroundai/playground-v2.5/blob/main/app.py) and [img2img-turbo](https://github.com/GaParmar/img2img-turbo/blob/main/gradio_sketch2image.py), respectively. Our safety checker is borrowed from [hart](https://github.com/mit-han-lab/hart).

Nunchaku is also inspired by many open-source libraries, including (but not limited to) [TensorRT-LLM](https://github.com/NVIDIA/TensorRT-LLM), [vLLM](https://github.com/vllm-project/vllm), [QServe](https://github.com/mit-han-lab/qserve), [AWQ](https://github.com/mit-han-lab/llm-awq), [FlashAttention-2](https://github.com/Dao-AILab/flash-attention), and [Atom](https://github.com/efeslab/Atom).