GitLab

Keep relevant weights in fp32 when `model._keep_in_fp32_modules` is set even when `accelerate` is not installed (#26225)

* fix bug where weight would not be kept in fp32

* nit

* address review comments

* fix test