mit-han-lab/svdq-int4-flux.1-schnell

Quantization Library: DeepCompressor   Inference Engine: Nunchaku

[Paper]  [Code]  [Website]  [Blog]

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.

Method

Quantization Method -- SVDQuant

Overview of SVDQuant. Stage1: Originally, both the activation X and weights W contain outliers, making 4-bit quantization challenging. Stage 2: We migrate the outliers from activations to weights, resulting in the updated activation and weight. While the activation becomes easier to quantize, the weight now becomes more difficult. Stage 3: SVDQuant further decomposes the weight into a low-rank component and a residual with SVD. Thus, the quantization difficulty is alleviated by the low-rank branch, which runs at 16-bit precision.

Nunchaku Engine Design

(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.

Model Description

• Developed by: MIT, NVIDIA, CMU, Princeton, UC Berkeley, SJTU and Pika Labs
• Model type: INT W4A4 model
• Model size: 6.64GB
• Model resolution: The number of pixels need to be a multiple of 65,536.
• License: Apache-2.0

Usage

Diffusers

Please follow the instructions in mit-han-lab/nunchaku to set up the environment. Then you can run the model with

import torch
from diffusers import FluxPipeline

from nunchaku.models.transformer_flux import NunchakuFluxTransformer2dModel

transformer = NunchakuFluxTransformer2dModel.from_pretrained("mit-han-lab/svdq-int4-flux.1-schnell")
pipeline = FluxPipeline.from_pretrained(
    "black-forest-labs/FLUX.1-schnell", transformer=transformer, torch_dtype=torch.bfloat16
).to("cuda")
image = pipeline("A cat holding a sign that says hello world", num_inference_steps=4, guidance_scale=0).images[0]
image.save("example.png")

Comfy UI

Please check comfyui/README.md for the usage.

Limitations

• The model is only runnable on NVIDIA GPUs with architectures sm_86 (Ampere: RTX 3090, A6000), sm_89 (Ada: RTX 4090), and sm_80 (A100). See this issue for more details.
• You may observe some slight differences from the BF16 models in details.

Citation

If you find this model useful or relevant to your research, please cite

@article{
  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},
  journal={arXiv preprint arXiv:2411.05007},
  year={2024}
}