fix bug:divide by zero in _maybe_log_save_evaluate() (#28251)

Co-authored-by: liujizhong1 <liujizhong1@xiaomi.com>

src/transformers/trainer.py

@@ -2238,7 +2238,7 @@ class Trainer:
             )
 
     def _maybe_log_save_evaluate(self, tr_loss, model, trial, epoch, ignore_keys_for_eval):
-        if self.control.should_log:
+        if self.control.should_log and self.state.global_step > self._globalstep_last_logged:
             if is_torch_tpu_available():
                 xm.mark_step()