Reproduce “PTQ4DiT: Post-training Quantization for Diffusion Transformers” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Junyi Wu, Haoxuan Wang, Yuzhang Shang, Mubarak Shah, Yan Yan
Researchers at the University of Illinois Chicago and the University of Central Florida present PTQ4DiT, the first dedicated post-training quantization method for Diffusion Transformers. It effectively quantizes DiTs to 8-bit precision with minimal quality loss and uniquely enables effective 4-bit weight quantization, achieving an FID of 7.09 for DiT-XL/2 on ImageNet 256x256 compared to 15.31 for previous methods.
Key insights
- Extreme magnitude values in DiT linear layers exhibit a "complementarity property" where they do not simultaneously coincide in the same channel of activations and weights, enabling their redistribution.
- Quantization strategies for Diffusion Transformers must account for the significant temporal variation of salient activation magnitudes across different denoising timesteps.
- Weighting the calibration process towards timesteps where the complementarity property is more pronounced can lead to greater salience reduction and improved quantization effectiveness.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.