Hey everyone - I had a quick question about Peft Models and using them with the trainer. If I use load_best_model_at_end=True in the Trainer arguments when training a peft model. When I call model.push_to_hub(...) I know the best adapter checkpoint is pushed up, but what happens if i call model.merge_and_unload().push_to_hub('my_merged_model_path') - will the default behaviour be that the best adapter checkpoint is loaded and merged and then the model is pushed, or will it just blindly merge the last adapter into the model and push it all up (i.e. forgetting the best adapter checkpoint altogether).
I’m trying to test it out currently but I don’t have an easy example to hand - dont suppose anyone has any insight?