Layerwise Error Attribution for Fast and Robust Mixed-Precision Post-Training Quantization
Layerwise error scores speed mixed-precision quantization up to 2,570x and resist corrupted calibration data.
The paper derives a layerwise probabilistic split of quantization error into propagated error and a local perturbation, then uses that local term as a separable score for mixed-precision bit allocation without an external solver. On DRUNet denoising with an average budget of 4 bits per weight, the method matches or improves mixed-precision baselines under clean calibration and gains up to 7.5 dB PSNR when calibration data are corrupted. Allocation is 28x to 2,570x faster than the studied baselines. A direct application of the same framework also improves state-of-the-art quantized diffusion models.
- A separable layerwise score allocates bits without an external combinatorial solver.
- DRUNet at 4 bits average matches baselines and gains up to 7.5 dB PSNR under corruption.
- Bit allocation is 28x to 2,570x faster than the studied mixed-precision baselines.
- A direct application also improves quantized diffusion model quality.
- AI models
- DRUNet
From the original article
1 min read at arxiv.orgMixed-precision post-training quantization is a network compression method that assigns bits layer by layer, under a global memory budget using a small calibration set. The main difficulties are to overcome the combinatorial nature of the allocation problem and to manage the sensitivity to small, potentially corrupted databases.
By Samy Houache · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.