Late Breaking Results: Quamba-SE: Soft-edge Quantizer for Activations in State Space Models
Fuente:
arXiv
Guardado en:
| Autores principales: | , |
|---|---|
| Formato: | Preprint |
| Publicado: |
2026
|
| Materias: | |
| Acceso en línea: | |
| Etiquetas: |
Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
|
| _version_ | 1866909990345768960 |
|---|---|
| author | Chen, Yizhi Hemani, Ahmed |
| author_facet | Chen, Yizhi Hemani, Ahmed |
| contents | We propose Quamba-SE, a soft-edge quantizer for State Space Model (SSM) activation quantization. Unlike existing methods, using standard INT8 operation, Quamba-SE employs three adaptive scales: high-precision for small values, standard scale for normal values, and low-precision for outliers. This preserves outlier information instead of hard clipping, while maintaining precision for other values. We evaluate on Mamba- 130M across 6 zero-shot benchmarks. Results show that Quamba- SE consistently outperforms Quamba, achieving up to +2.68% on individual benchmarks and up to +0.83% improvement in the average accuracy of 6 datasets. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2601_09451 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Late Breaking Results: Quamba-SE: Soft-edge Quantizer for Activations in State Space Models Chen, Yizhi Hemani, Ahmed Machine Learning Artificial Intelligence Hardware Architecture We propose Quamba-SE, a soft-edge quantizer for State Space Model (SSM) activation quantization. Unlike existing methods, using standard INT8 operation, Quamba-SE employs three adaptive scales: high-precision for small values, standard scale for normal values, and low-precision for outliers. This preserves outlier information instead of hard clipping, while maintaining precision for other values. We evaluate on Mamba- 130M across 6 zero-shot benchmarks. Results show that Quamba- SE consistently outperforms Quamba, achieving up to +2.68% on individual benchmarks and up to +0.83% improvement in the average accuracy of 6 datasets. |
| title | Late Breaking Results: Quamba-SE: Soft-edge Quantizer for Activations in State Space Models |
| topic | Machine Learning Artificial Intelligence Hardware Architecture |
| url | https://arxiv.org/abs/2601.09451 |