Skip to content

[docs] Add a note of _keep_in_fp32_modules #11851

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 3 commits into from
Jul 2, 2025
Merged

[docs] Add a note of _keep_in_fp32_modules #11851

Changes from 1 commit
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
8fbd332
update
a-r-r-o-w Jul 2, 2025
4acbf5c
Update docs/source/en/using-diffusers/schedulers.md
a-r-r-o-w Jul 2, 2025
e52d28d
Update schedulers.md
stevhliu Jul 2, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
3 changes: 3 additions & 0 deletions docs/source/en/using-diffusers/schedulers.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -242,3 +242,6 @@ unet = UNet2DConditionModel.from_pretrained(
)
unet.save_pretrained("./local-unet", variant="non_ema")
```

> [!TIP]
> The `dtype` in which a model is loaded can be specified with the `torch_dtype` argument when loading models with `from_pretrained`. On-the-fly dtype conversion can also be done using the Pytorch provided `.to()` method. An important distinction to keep in mind is that the latter converts all weights to the specified dtype, while the former takes into account a special model attribute (`_keep_in_fp32_modules`) when loading the weights. This is important in cases where some layers in the model must remain in FP32 precision for numerical stability and best generation quality. An example can be found [here](https://github.com/huggingface/diffusers/blob/f864a9a352fa4a220d860bfdd1782e3e5af96382/src/diffusers/models/transformers/transformer_wan.py#L374).