fix FA2 when using quantization

[pacman100]

What does this PR do?

1. when I use QLoRA+Flash Attention with bf16, I get the following warning of casting to float16 which is incorrect as it should be casting to bf16:

The input hidden states seems to be silently casted in float32, this might be related to the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in torch.float16.

This PR resolves this issue.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[younesbelkada]

Makes sense to me, we should first check if autocast is enabled even with quantization, then fallback to the quantization original dtype. Thanks for the fix @pacman100 !

[amyeroberts]

@pacman100 Thanks for fixing this!

[susnato]

Hi, this behavior of casting to float16 is still present in these models - whisper, bart, phi, distilbert...I will create a PR to fix it.