fix single (#5162)

PaddleNLP/benchmark/transformer/static/train.py

@@ -119,10 +119,12 @@ def do_train(args):
         batch_id = 0
         batch_start = time.time()
         pass_start_time = batch_start
-        for data in train_loader():
+        for data in train_loader:
             # NOTE: used for benchmark and use None as default.
             if args.max_iter and step_idx == args.max_iter:
                 return
+            if trainer_count == 1:
+                data = [data]
             train_reader_cost = time.time() - batch_start
 
             outs = exe.run(compiled_train_program,