Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2025
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2506.09932 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866915381346566144 |
|---|---|
| author | Federici, Marco Del Chiaro, Riccardo van Breugel, Boris Whatmough, Paul Nagel, Markus |
| author_facet | Federici, Marco Del Chiaro, Riccardo van Breugel, Boris Whatmough, Paul Nagel, Markus |
| contents | Diffusion models represent the cutting edge in image generation, but their high memory and computational demands hinder deployment on resource-constrained devices. Post-Training Quantization (PTQ) offers a promising solution by reducing the bitwidth of matrix operations. However, standard PTQ methods struggle with outliers, and achieving higher compression often requires transforming model weights and activations before quantization. In this work, we propose HadaNorm, a novel linear transformation that extends existing approaches by both normalizing channels activations and applying Hadamard transforms to effectively mitigate outliers and enable aggressive activation quantization. We demonstrate that HadaNorm consistently reduces quantization error across the various components of transformer blocks, outperforming state-of-the-art methods. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2506_09932 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | HadaNorm: Diffusion Transformer Quantization through Mean-Centered Transformations Federici, Marco Del Chiaro, Riccardo van Breugel, Boris Whatmough, Paul Nagel, Markus Computer Vision and Pattern Recognition Artificial Intelligence Diffusion models represent the cutting edge in image generation, but their high memory and computational demands hinder deployment on resource-constrained devices. Post-Training Quantization (PTQ) offers a promising solution by reducing the bitwidth of matrix operations. However, standard PTQ methods struggle with outliers, and achieving higher compression often requires transforming model weights and activations before quantization. In this work, we propose HadaNorm, a novel linear transformation that extends existing approaches by both normalizing channels activations and applying Hadamard transforms to effectively mitigate outliers and enable aggressive activation quantization. We demonstrate that HadaNorm consistently reduces quantization error across the various components of transformer blocks, outperforming state-of-the-art methods. |
| title | HadaNorm: Diffusion Transformer Quantization through Mean-Centered Transformations |
| topic | Computer Vision and Pattern Recognition Artificial Intelligence |
| url | https://arxiv.org/abs/2506.09932 |