KroQuant: Kronecker-Structured Block Transforms for Efficient Post-Training Quantization of Diffusion Transformers
A learned Kronecker transform lets 4-bit diffusion transformers avoid activation outliers without paying for full dense transforms.
KroQuant applies an invertible transform inside each 32-element activation block, then quantizes weights and activations to W4A4. The authors say its kernel is up to 14% faster than SmoothQuant on an MI350 GPU, while using less than half the parameters of per-channel scaling. With offline LoRaQ weight calibration, it produces outputs closer to the full-precision reference than SVDQuant and LoRaQ on PixArt-Sigma, SANA, and FLUX.1-schnell. ArXiv · AI/CL/LG's note
KroQuant applies an invertible transform inside each 32-element activation block, then quantizes weights and activations to W4A4. The authors say its kernel is up to 14% faster than SmoothQuant on an MI350 GPU, while using less than half the parameters of per-channel scaling. With offline LoRaQ weight calibration, it produces outputs closer to the full-precision reference than SVDQuant and LoRaQ on PixArt-Sigma, SANA, and FLUX.1-schnell. ArXiv · AI/CL/LG's note
score 5