Tail-Aware HiFloat4: W4A4 Post-Training Quantization for Wan2.2

In three linesTail-Aware HiFloat4 applies W4A4 post-training quantization to the Wan2.2 text-to-video generation model. The method adapts ViDiT-Q using HiFloat4 format, quantizes transformer linear layers, preserves numerically sensitive modules in high precision, and introduces activation-tail-aware percentile calibration to reduce impact of rare outliers.

Summary generated by Claude — human-verified