Attention Blocks and Model Loading

[AasherH]

I am trying to acquire the attentions from the pretrained model (rt_1_x_tf_trained_for_002272480_step). In pytorch with LLaVA, for instance, there is an option, e.g., model.output_attentions=True. Is there a similar option in this repo?

From what I could find from browsing the models subfolder, no such functionality exists. Does anyone know how I could go about acquiring the attention matrices for each block?

Moreover, what is the appropriate way to load the model to access its underlying architecture? I am used to keras, but the saved_model format prevents use of keras functionality which would make this task simpler.

Thank you for any information you can provide!