Stable and low-precision training for large-scale vision-language models

We introduce new methods for 1) accelerating and 2) stabilizing training for large language-vision models.