Method clip_gradient

trainers.py:397–399

Clip the gradient norm of the parameters of a non-FSDP policy.

(self)

    def clip_gradient(self):
        """Clip the gradient norm of the parameters of a non-FSDP policy."""
        return torch.nn.utils.clip_grad_norm_(self.policy.parameters(), self.config.max_grad_norm).item()

    def write_state_dict(self, step: int, state: Dict[str, torch.Tensor], metrics: Dict, filename: str, dir_name: Optional[str] = None):
        """Write a checkpoint to disk."""