Fix #7304 (#7305)

src/transformers/trainer.py

@@ -1334,9 +1334,9 @@ class Trainer:
         elif is_torch_tpu_available():
             # tpu-comment: Get all predictions and labels from all worker shards of eval dataset
             if preds is not None:
-                preds = nested_xla_mesh_reduce("eval_preds", preds)
+                preds = nested_xla_mesh_reduce(preds, "eval_preds")
             if label_ids is not None:
-                label_ids = nested_xla_mesh_reduce("eval_label_ids", label_ids, torch.cat)
+                label_ids = nested_xla_mesh_reduce(label_ids, "eval_label_ids")
             if eval_losses is not None:
                 eval_losses = xm.mesh_reduce("eval_losses", torch.tensor(eval_losses), torch.cat).tolist()