update create_model_card to properly save peft details when using Trainer with PEFT

[pacman100]

What does this PR do?

1. When using Trainer with PEFT, model.save_pretrained in PEFT adds PEFT-specific details to the existing model card or creates a new model card and adds these details. However, trainer.create_model_card is called after model is saved and it overwrites the entire file thereby nullifies the addition of details related to PEFT such as library name, quantization used and PEFT version.
2. This PR fixes the above issue and thereby adds backs the PEFT details to the model card. This will help in better organization and understanding usage of PEFT on Hub.
3. Example of the repo with the PR usage: https://huggingface.co/smangrul/mistral_lora_clm_with_added_tokens

[muellerzr]

Solution makes sense to me! Thanks!

[BenjaminBossan]

I'm not very knowledgeable about Trainer, so can't comment on the whole integration.

From my understanding, this seems a little bit hacky to me, because we first create the model card with PEFT, then completely overwrite it with Trainer, then re-add PEFT-related content. It feels like the proper solution would be for the Trainer to update the model card if it already exists. But I understand that this would be more work, so I'd be okay with this more hacky approach.

[pacman100]

From my understanding, this seems a little bit hacky to me, because we first create the model card with PEFT, then completely overwrite it with Trainer, then re-add PEFT-related content. It feels like the proper solution would be for the Trainer to update the model card if it already exists. But I understand that this would be more work, so I'd be okay with this more hacky approach.

Yes, Trainer should ideally update if README is already there, but it rewrites everything from the TrainerState and as such appending/updating would be trickier. Open to ideas on making this cleaner.

[BenjaminBossan]

Thanks for the update.

Yes, Trainer should ideally update if README is already there, but it rewrites everything from the TrainerState and as such appending/updating would be trickier. Open to ideas on making this cleaner.

I'm not very knowledgeable about Trainer, so I don't have any suggestion for a better solution.

[muellerzr]

I can't really see a better solution currently either, so what we have here works for now