Samin Aref
@samin.mas.to.ap.brid.gy
Researcher and educator
Faculty at the University of Toronto
Networks, Machine Learning, Operations Research, and (Social) Data Science (Reblog⇏Endorse)
[bridged from https://mas.to/@samin on the fediverse by https://fed.brid.gy/ ]
Faculty at the University of Toronto
Networks, Machine Learning, Operations Research, and (Social) Data Science (Reblog⇏Endorse)
[bridged from https://mas.to/@samin on the fediverse by https://fed.brid.gy/ ]
make them more resource efficient at the cost of potential performance reduction. In a new preprint, we extended an existing frontier method (ApiQ) for quantizing weights of LLMs down to 1 or 2 bits. On LLaMA-2 models, our method covers 7.5-10.8% of the accuracy gap between the full-precision […]
Original post on mas.to
mas.to
April 24, 2025 at 11:14 PM
make them more resource efficient at the cost of potential performance reduction. In a new preprint, we extended an existing frontier method (ApiQ) for quantizing weights of LLMs down to 1 or 2 bits. On LLaMA-2 models, our method covers 7.5-10.8% of the accuracy gap between the full-precision […]