Dreambooth: fix #1566: Use autocast to correct full vs half precision error

[DrewWalkup]

To address Dreambooth: fix #1566, a potential solution using autocast.

Further documentation here: https://discuss.pytorch.org/t/training-with-half-precision/11815/17

Using suggested implementation from here: https://discuss.pytorch.org/t/runtimeerror-input-type-torch-cuda-floattensor-and-weight-type-torch-halftensor-should-be-the-same/104312/6

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint.

[patrickvonplaten]

I think the problem was fixed in #1567 . Also, we should not use autocast as explained here: #511

cc @pcuenca

[pcuenca]

I agree, using autocast in combination with accelerate can create all sort of issues. @dnwalkup did you try after #1567 was merged? I think it should be working fine now.

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.