trl
trl copied to clipboard
Neftune is applied twice; in trl and transformers BOTH!
As long as I read is right, neftune is being applied twice both in trl and transformers.
https://github.com/huggingface/trl/blob/39a7d1c121d26224fd7455d3d2038e0d20831c54/trl/trainer/sft_trainer.py#L440 calls Trainer.train
and Trainer.train have self.model = self._activate_neftune(self.model)
inside.
Please let me know if I misunderstood.
Releated PR
- https://github.com/huggingface/trl/pull/871