Adapting a HookedTransformer to a model that is not part of the existing models

[Noam-Diamant]

Is there a way to create a HookedTransformer.from_pretrained method for a model that is not in the Model Properties Table of transformerlens?
I need to use HookedTransformer for my research, which involves working with the CLIP language model, and for that I would like to create a HookedTransformer on its language model. Is there a way to use the library so that I can create a HookedTransformer for a model that is not in the existing list of models?

[bryce13950]

At this moment, there is not a simple way to do that, unless the model is compatible with an existing architecture. This will come in the future, but if you want to share a timeline with me, I may be able to get it into the current batch of models that are being added.

[jlarson4]

Hi @Noam-Diamant! If there is a model you're still looking to study, please take a look at the recent released TransformerLens 3.0! It includes support for 9,000 models, including a selection of LLaVA models that use a CLIP vision encoder. If you need support for a specific CLIP architecture, let me know and we can work on creating an architecture adapter.