Module-wise Quantization for Transformer-based Neural Video Coding

Transformer-based neural video codecs achieve strong rate-distortion performance, but their high complexity and reliance on floating-point arithmetic hinder practical deployment. In this work, we propose a module-wise post-training quantization (PTQ) pipeline for the MaskCRT codec, where each module is wrapped to interface with a quantization framework, enabling input tracing, calibration, and export of reusable encodings without retraining or modifying the codec core. Experimental results show that moderate precisions (12-16 bits) preserve near-anchor performance, whereas 8-bit activations cause severe degradation and unstable rate-distortion behavior. Single-module analysis identifies Conditional Motion and Residual as the most precision-sensitive components. Guided by this analysis, an RD-BOP-guided allocation policy greedily assigns bit-widths to maximize rate-distortion recovery per unit of BOP cost. At the same BOP budget as uniform 12-bit integer precision (W12A12), the proposed policy reduces aggregate BD-rate degradation from +3.59% to +2.20% over HEVC-B and UVG. To the best of our knowledge, this is the first systematic PTQ study for a Transformer-based neural video codec.