Fix LoRA loading for flattened CLIP text encoders

[BeyondBirthday07]

What does this PR do?

Fixes #14860

CLIPTextModel is flattened in newer transformers versions, so its module names no longer start with text_model.. Older LoRA checkpoints can still contain this prefix.

Because of that, the text encoder rank lookup does not find the LoRA keys, leaving rank empty and causing an IndexError.

This strips the text_model. prefix when loading into a flattened text encoder.

Added a regression test for this case. I also tested the fix with the LoRA checkpoint shared in #14860.

No new dependencies.

Tests

python -m pytest tests/lora/test_peft_utils.py tests/lora/test_lora_loader_utils.py -q
python -m ruff check src/diffusers/loaders/lora_base.py tests/lora/test_lora_loader_utils.py
python -m ruff format --check src/diffusers/loaders/lora_base.py tests/lora/test_lora_loader_utils.py

Before submitting

• Did you use an AI agent (Claude Code, Codex, Cursor, etc.) to help with this PR? If so:
  • Did you read the Coding with AI agents guide?
  • Did you run the self-review skill on the diff?
  • Did you share the final self-review notes in the PR description or a comment?
• Did you read the contributor guideline?
• Did you read our philosophy doc? (important for complex PRs)
• Was this discussed/approved via a GitHub issue or the forum? Please add a link to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?
• Are you the author (or part of the team) of the model/pipeline (only applicable for model/pipeline related PRs)?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

@sayakpaul @BenjaminBossan

[BenjaminBossan]

I guess this should work but the fix is very specific, I wonder if there is a more general solution.

[sayakpaul]

But this is also a very transformers-specific quirk, IMO. So, it seems okay to me. Do you have any ideas?

[BeyondBirthday07]

Yeah i thought about making it more general too, but this looks pretty specific to the transformers CLIPTextModel flattening, so I kept the fix narrow.

[BenjaminBossan]

IMO, a test should be closer to how a user would load such a checkpoint via a pipeline, using the mixin class directly is not normally done. If the maintainers are fine with this, I won't object though.

[sayakpaul]

Yes, I agree with you, @BenjaminBossan. However, internally, the call gets dispatched to load_lora_into_text_encoder from load_lora_weights(). So, seems okay to me as a minimal and self-contained use case.

[BeyondBirthday07]

Yeah, I kept the test at the loader level since load_lora_weights() ends up calling the same path anyway.