Revert "Enforce same behavior as PyTorch 2.0 for older versions" (#22163)

Revert "Enforce same behavior as PyTorch 2.0 for older versions (#22136)"

This reverts commit 1c801d65.

src/transformers/trainer.py

@@ -1813,7 +1813,7 @@ class Trainer:
         # _total_loss_scalar is updated everytime .item() has to be called on tr_loss and stores the sum of all losses
         self._total_loss_scalar = 0.0
         self._globalstep_last_logged = self.state.global_step
-        model.zero_grad(set_to_none=True)
+        model.zero_grad()
 
         self.control = self.callback_handler.on_train_begin(args, self.state, self.control)
 
@@ -1971,7 +1971,7 @@ class Trainer:
                     if optimizer_was_run and not self.deepspeed:
                         self.lr_scheduler.step()
 
-                    model.zero_grad(set_to_none=True)
+                    model.zero_grad()
                     self.state.global_step += 1
                     self.state.epoch = epoch + (step + 1 + steps_skipped) / steps_in_epoch
                     self.control = self.callback_handler.on_step_end(args, self.state, self.control)