Tail-Aware HiFloat4: W4A4 Post-Training Quantization for Wan2.2
This work addresses the challenge of low-bit (W4A4) post-training quantization for text-to-video generation models, specifically for the Wan2.2 architecture, offering a practical solution to reduce memory and compute costs.
The authors adapted the ViDiT-Q post-training quantization pipeline to Wan2.2 using HiFloat4 format, achieving W4A4 quantization of main linear layers while preserving high precision for sensitive boundary modules. Their tail-aware percentile calibration module reduces outlier influence, enabling competitive text-to-video generation quality with low-bit quantization.
This report describes Tail-Aware HiFloat4, our submission to the low-bit text-to-video generation quantization challenge. Our method adapts the public ViDiT-Q post-training quantization pipeline to Wan2.2 under the HiFloat4 numerical format. We quantize the main linear layers in both Wan2.2 transformer modules with W4A4 HiFloat4 fake quantization, keep numerically sensitive boundary modules in high precision, and introduce an activation-tail-aware percentile calibration module for channel-mask construction. Together with compact PTQ-state restoration, this design reduces the influence of rare calibration outliers while keeping the runtime HiFloat4 arithmetic and sampling pipeline unchanged.