fixed the issue of DPO trainer that using one node and mutiple GPUs and set...

fixed the issue of DPO trainer that using one node and mutiple GPUs and set the device_map='auto' (#29695)

* fixed the issue of DPO trainer that using one node and mutiple GPUs

* before update, add the assert

* run the ruff formatter

* Update src/transformers/trainer.py

Thank you.
Co-authored-by: Younes Belkada <49240599+younesbelkada@users.noreply.github.com>

* remember to do make style and make quality before commit

* Update src/transformers/trainer.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

---------
Co-authored-by: Younes Belkada <49240599+younesbelkada@users.noreply.github.com>
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

src/transformers/trainer.py

@@ -2124,6 +2124,10 @@ class Trainer:
                     # if loss is nan or inf simply add the average of previous logged losses
                     tr_loss += tr_loss / (1 + self.state.global_step - self._globalstep_last_logged)
                 else:
+                    if tr_loss.device != tr_loss_step.device:
+                        raise ValueError(
+                            f"Calculated loss must be on the original device: {tr_loss.device} but device in use is {tr_loss_step.device}"
+                        )
                     tr_loss += tr_loss_step
 
                 self.current_flos += float(self.floating_point_ops(inputs))