Specify gradients in model preparation

[muellerzr]

Adjusts the docstring to specify that only models whose gradients will be utilized should be passed through accelerator.prepare

[sgugger]

That's actually not accurate. The model is also prepared to handle mixed precision training in the prepare_model method, so any model should be sent to the prepare call ideally.

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[sgugger]

Thanks!