Fix float precision in Finetune CLI

facebookresearch / seamless_communication

Foundational Models for State-of-the-Art Speech and Text Translation

Other

10.94k stars 1.06k forks source link

Fix float precision in Finetune CLI #399

Closed zrthxn closed 7 months ago

zrthxn commented 7 months ago

The finetune CLI would load the model in float16 and then attempt to train. Doing it this was, the model would unable to converge. Here the model is loaded in float32 and then we allow autocasting where it is needed.