HyGenQ Improves Quantization for Hybrid Generative Models

Researchers have developed HyGenQ, a new post-training quantization (PTQ) framework specifically tailored for hybrid iterative generative models (IGMs). While PTQ is crucial for accelerating the inference of these models, directly applying standard PTQ methods often leads to model collapse due to two main issues: excessive outliers (EOs) in activations and amplified anomalies (AAs) from minor quantization errors. EOs create a trade-off between precision and coverage, while AAs cause a mismatch between calibration and inference, leading to iterative degradation. HyGenQ addresses these challenges through two key components: Hierarchical Cluster Decoupling (HCD) and Scaling Recalibration (SR). HCD identifies and isolates outlier channels using a multi-stage clustering process, preserving normal value precision while handling EOs. SR scales AAs beyond Gaussian bounds, preventing model collapse caused by aggressive truncation. Extensive experiments demonstrate that HyGenQ successfully quantizes various hybrid IGMs to 8-bit precision (W8A8), significantly outperforming existing baselines and validating its robustness across different model families.