Fix eval_accumulation_steps leading to incorrect metrics (#24756)

Fix eval steps

src/transformers/trainer.py

@@ -3154,7 +3154,7 @@ class Trainer:
             self.control = self.callback_handler.on_prediction_step(args, self.state, self.control)
 
             # Gather all tensors and put them back on the CPU if we have done enough accumulation steps.
-            if args.eval_accumulation_steps is not None and (step + 1) % args.eval_accumulation_steps == 0:
+            if args.eval_accumulation_steps is not None and self.accelerator.sync_gradients:
                 if losses_host is not None:
                     losses = nested_numpify(losses_host)
                     all_losses = losses if all_losses is None else np.concatenate((all_losses, losses), axis=0)