Adapting replit transformers support for training

I am trying to finetune replit code model (replit/replit-code-v1-3b 路 Hugging Face) using the same procedure in alpaca-lora, aiming to make the model instruction following.
However, when I try to use the model with transformers.Trainer, it turns out the replit model forward function does not support labels and input_embeds which I guess are used by Trainer to update the weights: ReplitLM/ at 1a421995c10c880379f52e26b5ba495c87bac3e2 路 replit/ReplitLM 路 GitHub
I鈥檓 wondering if there鈥檚 a guide or specific steps I need to follow in order to modify the replit model code base so that it supports training with Trainer class