一种适用于自回归掩码生成模型的后训练量化方法
By employing hierarchical clustering decoupling and scaling recalibration mechanisms, the problems of extreme outliers and anomalous amplification in the MAR model are solved, achieving stable deployment under low bit precision, reducing computation and storage costs, and making it suitable for edge devices.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- BEIJING JIAOTONG UNIV
- Filing Date
- 2025-09-15
- Publication Date
- 2026-07-17
AI Technical Summary
Existing post-training quantization methods cannot effectively handle extreme outliers and anomalous amplification in autoregressive mask generation models (MAR), resulting in unstable model deployment, excessively high computational and storage costs, and difficulty in application in resource-constrained environments.
Hierarchical clustering decoupling mechanism (HCD) and scaling recalibration mechanism (SR) are adopted to identify and decouple abnormal channels, dynamically adjust the quantization range, and cooperate with the full-process static quantization strategy to achieve stable deployment under low bit precision.
While maintaining generation quality, it significantly reduces storage and computing costs, enabling stable inference deployment of MAR models at low bit precision, suitable for edge devices and resource-constrained environments.
Smart Images
Figure CN121168554B_ABST