Layerwise Error Attribution for Fast and Robust Mixed-Precision Post-Training Quantization
The method uses a layerwise probabilistic analysis of quantization error to create a fast and robust mixed-precision post-training quantization algorithm. It separates propagated error from local layer perturbation, allowing for quick allocation decisions without external solvers. This approach achieves state-of-the-art performance while significantly improving robustness against corrupted calibration data.