Fix: Prevent crash when trying to load SigLIP/CLIP encoders with transformers >= 5.6.0

[Aaron-d-k]

Description

Currently, trying to run any model using the SigLIP or CLIP encoder (Eg: google/gemma-3-4b-it) using a default installation of transformer-lens from PyPI crashes due to a change introduced in the transformers library in 5.6.0 (commits: huggingface/transformers#44431). Both SiglipVisionModel and CLIPVisionModel had their structure slightly changed.
This issue arises because currently pyproject.toml in transformer-lens only specifies a minimum version.
There are three fixes I considered:

1. Adding <5.6.0 to the toml
2. Porting the entire library to 5.6.0
3. Adding a backward compatible method that checks for the version and corrects for it.

I have done 3. My change should not affect any users with transformers<5.6.0 or those who do not use the model, so is entirely backward compatible. It basically checks if the required attribute is missing when set_original_component is called and tries to monkey patch it in if it is necessary.

Type of change

• Bug fix (non-breaking change which fixes an issue)

Screenshots

Before (transformers==5.8.1, demos/Gemma3_Multimodal.ipynb):

After fix:

Checklist:

• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works (I am unsure how I should add tests for this)
• New and existing unit tests pass locally with my changes
• I have not rewritten tests relating to key interfaces which would affect backward compatibility

[jlarson4]

Hi @Aaron-d-k! Thank you for both finding and fixing this bug. This is an excellent solution to the problem.

For testing, I'd recommend following the example set in tests/unit/model_bridge/supported_architectures/test_gemma3_multimodal_adapter.py. It build a stub using SimpleNamespace that doesn't download the model, but can validate shape.