[fix] CLIPTextModel with transformers >= 5.6 and from_single_file

[asomoza]

What does this PR do?

This PR fixes the issue when using recent transformers versions and from_single_file with CLIPTextModel . This prevents apps and libraries to pin transformers to more recent versions when they supports this.

Tested with an older version of transformers without the fix and with a newer one (latest as today) with the fix, using SD1.5 and SD2.1 and the images are identical.

SD 1.5

older without fix	newer with fix

SD 2.1

older without fix	newer with fix

fixes #13833

Who can review?

@kappacommit

[asomoza]

@askserge review this PR

[github-actions]

🤗 Serge says:

Clean, correct fix for CLIPTextModel compatibility with transformers ≥5.6 where the text_model wrapper was flattened.

Correctness

• The hasattr(model, "text_model") check correctly distinguishes between the old (CLIPTextModel with wrapper) and new (flattened) layouts, while leaving CLIPTextModelWithProjection (which still has text_model) unaffected.
• Both convert_ldm_clip_checkpoint and convert_open_clip_checkpoint produce keys prefixed with text_model. in the diffusers format. The removeprefix("text_model.") call correctly strips this for the flattened model, and is a no-op for keys that don't carry the prefix (e.g. text_projection.weight).
• removeprefix (Python 3.9+) is already used elsewhere in the codebase.

No issues found. LGTM.

13 LLM turns · 14 tool calls · 66.3s · 148568 in / 2630 out tokens

[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.

[ghunkins]

Thanks for the work @asomoza, would love to see this on main 🙏