[fix] Diffusers dtype in code snippets

[stevhliu]

Use torch_dtype instead of dtype (not supported by from_pretrained) for the Diffusers Hub snippets (see Slack thread for more context)

cc @sayakpaul @pagezyhf

---

Note

Low Risk
Documentation-only string changes in snippet templates; no runtime or API behavior in the app.

Overview
Updates generated Diffusers Hub code snippets so from_pretrained calls pass torch_dtype instead of dtype, which Diffusers does not accept on that API.

The change is applied consistently across text-to-image, img2img, video, LoRA, textual inversion, Flux fill, and inpainting snippet templates in model-libraries-snippets.ts, so copy-paste examples load pipelines with the intended half/bfloat precision.

^{Reviewed by Cursor Bugbot for commit f715770. Bugbot is set up for automated code reviews on this repo. Configure here.}

[julien-c]

there's no way to alias dtype to torch_dtype for backward compatibility?

[stevhliu]

yeah i think we can also do that and it'd be aligned with transformers as well :) wdyt @sayakpaul?

[sayakpaul]

It is currently not planned.

[julien-c]

let's add the alias before merging this then. Should be simple

[pagezyhf]

hello, are there any chance to have this one merged by the 15th for a partner launch?

[stevhliu]

@julien-c, alias has been added in huggingface/diffusers#14162 so this should be ready to merge

[julien-c]

so we merge this and in a couple of weeks (when people upgrade their diffusers version) we revert it, is this correct?

[stevhliu]

yeah i believe so, diffusers is planning on deprecating torch_dtype - see huggingface/diffusers#14162 (comment) - and consolidate to dtype

[julien-c]

let's remember to revert in a couple of weeks