fix bf16 model type issue for ia3

[sywangyi]

No description provided.

[sywangyi]

@BenjaminBossan please help review.

[sywangyi]

if set torch_dtype is bfloat16 in model.from_pretained(), all adapter layers after get_peft_model(xxx) should be bfloat16 as well. fix issue in ia3 and adalora, also ia3 forward return dtype should be aligned with the base_layer.forward() output.

[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.

[BenjaminBossan]

Thanks, those changes to correctly enable bf16 look good. I'll request another review by the co-maintainers just to be sure.

[BenjaminBossan]

@pacman100 Could you please check again?

[pacman100]

Thank you, @sywangyi, for the fixes!