Fix the behavior of collecting 'num_input_tokens_seen' (huggingface#2…

…9099)

fix the behavior of collecting 'num_input_tokens_seen'

See huggingface#28791 for more details.

src/transformers/trainer.py

@@ -2097,7 +2097,12 @@ def _inner_training_loop(
                             "a `main_input_name` attribute to the model class you are using."
                         )
                     else:
-                        self.state.num_input_tokens_seen += self.accelerator.gather(inputs[main_input_name]).numel()
+                        input_device = inputs[main_input_name].device
+                        self.state.num_input_tokens_seen += torch.sum(
+                            self.accelerator.gather(
+                                torch.tensor(inputs[main_input_name].numel(), device=input_device, dtype=torch.int64)
+                            )
+                        ).item()
                 if rng_to_sync:
                     self._load_rng_state(resume_from_checkpoint)
                     rng_to_sync = False