Stable and low-precision training for large-scale vision-language models Mitchell Wortsman 1 Tim Dettmers 1 Luke Zettlemoyer