RaZeR: Pushing the Limits of NVFP4 Quantization with Redundant Zero Remapping
Yuzong Chen, Xilai Dai, Jake Hyun, Chi-Chih Chang, Wonsuk Jang, Yuheng Wu, Thierry Tambe, Jae-sun Seo, Mohamed S. Abdelfattah · 2025
Read RaZeR: Pushing the Limits of NVFP4 Quantization with Redundant Zero Remapping by Yuzong Chen, Xilai Dai, Jake Hyun, Chi-Chih Chang, Wonsuk Jang, Yuheng Wu, Thierry Tambe, Jae-sun Seo, Mohamed S. Abdelfattah (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-04052.txt.
Description
Yuzong Chen and co-authors squeeze more value from a compact number format for large language models, in this 2025 paper. The format, NVFP4, stores model weights in just four bits, which saves memory and speeds answers. The authors spotted waste in how those bits are spent. The format holds both a plus zero and a minus zero, though one would do, and its scaling values carry a sign bit that is never used.
Their fix, called RaZeR, remaps the spare zero to hold an extra value instead. They also note that model weights tolerate cruder scaling than the format gives. Together the changes sharpen accuracy without growing the storage size. The work targets cheaper, faster serving of large models on current hardware.
Catalog record
| CREATOR | Yuzong Chen, Xilai Dai, Jake Hyun, Chi-Chih Chang, Wonsuk Jang, Yuheng Wu, Thierry Tambe, Jae-sun Seo, Mohamed S. Abdelfattah |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 93 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2501-04052