Reset loss to zero on logging in Trainer to avoid bfloat16 issues (#8561)

* make tr_loss regular float

* Revert "make tr_loss regular float"

This reverts commit c9d7ccfaf0c4387187b0841694f01ec0ffd5f4ba.

* reset loss at each logging step

* keep track of total loss with _total_loss_scalar

* add remaining tr_loss at the end

src/transformers/trainer.py

@@ -696,8 +696,10 @@ class Trainer:
         self.state.is_local_process_zero = self.is_local_process_zero()
         self.state.is_world_process_zero = self.is_world_process_zero()
 
+        # tr_loss is a tensor to avoid synchronization of TPUs through .item()
         tr_loss = torch.tensor(0.0).to(self.args.device)
-        self._logging_loss_scalar = 0
+        # _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 = 0
         self._total_flos = self.state.total_flos
         model.zero_grad()
@@ -812,23 +814,26 @@ class Trainer:
             self.log({"total_flos": self.state.total_flos})
 
         self.control = self.callback_handler.on_train_end(self.args, self.state, self.control)
+        # add remaining tr_loss
+        self._total_loss_scalar += tr_loss.item()
 
-        return TrainOutput(self.state.global_step, tr_loss.item() / self.state.global_step)
+        return TrainOutput(self.state.global_step, self._total_loss_scalar / self.state.global_step)
 
     def _maybe_log_save_evaluate(self, tr_loss, model, trial, epoch):
         if self.control.should_log:
             logs: Dict[str, float] = {}
             tr_loss_scalar = tr_loss.item()
-            logs["loss"] = (tr_loss_scalar - self._logging_loss_scalar) / (
-                self.state.global_step - self._globalstep_last_logged
-            )
+            # reset tr_loss to zero
+            tr_loss -= tr_loss
+
+            logs["loss"] = tr_loss_scalar / (self.state.global_step - self._globalstep_last_logged)
             # backward compatibility for pytorch schedulers
             logs["learning_rate"] = (
                 self.lr_scheduler.get_last_lr()[0]
                 if version.parse(torch.__version__) >= version.parse("1.4")
                 else self.lr_scheduler.get_lr()[0]
             )
-            self._logging_loss_scalar = tr_loss_scalar
+            self._total_loss_scalar += tr_loss_scalar
             self._globalstep_last_logged = self.state.global_step
 
             self.log(logs)