QDiTx: Arbitrary Bit-Width Quantization Framework for Diffusion Transformer
Abstract
QDiTx is an arbitrary-bit-width weight-quantization framework for Diffusion Transformers. It addresses accumulated quantization error during iterative denoising and GPU memory-access inefficiencies caused by non-power-of-two bit widths through resolution-aware precision search, cascaded quantization, warp-aware weight layout, and parallel dequantization. In weight-loading-bound settings, QDiTx achieves up to 7× speedup while maintaining generation quality close to FP32.
Type
Publication
In IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
Qianlong Sang
Fifth-Year Computer Science
Ph.D. Student
My research focuses on operating systems, mobile and edge systems, and efficient on-device AI inference.