How to load after calling trainer.model.push_to_hub() on a fine tuned model?

I fine-tuned a falcon-7b model and called trainer.train() using SFTTrainer from huggingface’s trl package.

Calling trainer.model.push_to_hub('hub_name') pushes three files to the hugginface repository–a, an adapter_config.json, and adapter_model.bin. I’m having trouble finding any documentation that describes how to use these file formats.

At first I found this huggingface page, “Using Adapter Transformers at Hugging Face.)”, and it says to use the class “AutoModelWithHeads”. However, this class won’t load after downloading the adapter-transforms pip package on Google Colab, I think it’s deprecated? Also, calling

adapter_name = model.load_adapter(ADAPTER)
model.active_adapters = adapter_name

with AutoModel results in this error:

AttributeError: ‘RWForCausalLM’ object has no attribute ‘load_adapter’

How can you apply adapter_model.bin files? My code was based on the official Falcon-7b huggingface tutorial, where they link a Google Colab notebook on that page.

Same Problem here with LLama 2, did you fix this somehow or did you find a solution?