(Maths-IA) Layerwise Error Attribution for Robust Mixed-Precision Post-Training Quantization

Mixed-precision post-training quantization assigns different bit-widths to different blocks of a network. The challenge is to estimate block sensitivity from a small calibration set without testing every possible allocation. In this talk, we present a layerwise analysis of quantization error that separates the error propagated from previous layers from the local perturbation introduced by the current layer. This local term motivates a practical score computed for each block and candidate bit-width using full-precision activations. Because the scores are independent of the choices made for other blocks, they can be computed once and reused by a simple greedy allocation algorithm. We evaluate this approach on DRUNet image denoising under clean and corrupted calibration data, and on a CelebA-HQ latent diffusion model. The method achieves competitive or better image quality at moderate bit-widths, remains stable under the tested calibration corruptions, and reduces allocation time from minutes or hours to seconds in the reported experiments.