andsteing / aqt

AQT : Accurate Quantized Training

AQT is a quantization library designed to allow utilization of low-bit and high-performance numerics of contemporary ML hardware accelerators. AQT supports both research and production¹, but focuses on the latter.

Citing AQT

Please use a following bibtex entry:

@software{aqt2022github,
  author = {Lew, Lukasz and Feinberg, Vlad and Agrawal, Shivani and Lee, Jihwan and Malmaud, Jonathan and Wang, Lisa and  Dormiani, Pouya and Pope, Reiner },
  title = {AQT: Accurate Quantized Training)},
  url = {http://github.com/google/aqt},
  year = {2022},
}

Footnotes

1. The support for research is exemplified by having a state of the art quantization quality on standard models such as ResNet and Transformer. The production aspect is defined as high performance and robust out-of-the-box working results with good defaults. ↩