Trainer Loads old TrainingArguments

I am trying to resume training from a checkpoint. But then the trainer ignores my current training arguments, and instead train with the training arguments that is stored with the checkpoint. I am using SFT Trainer.