diff --git a/benchmark/transformer/static/predict.py b/benchmark/transformer/static/predict.py
index 245e690e2d0f1..efe25b6209ced 100644
--- a/benchmark/transformer/static/predict.py
+++ b/benchmark/transformer/static/predict.py
@@ -20,6 +20,7 @@
 logging.basicConfig(level=logging.INFO, format=FORMAT)
 logger = logging.getLogger(__name__)
 
+
 def cast_parameters_to_fp32(place, program, scope=None):
     all_parameters = []
     for block in program.blocks:
@@ -33,6 +34,7 @@ def cast_parameters_to_fp32(place, program, scope=None):
             data = np.array(tensor)
             tensor.set(np.float32(data), place)
 
+
 def parse_args():
     parser = argparse.ArgumentParser()
     parser.add_argument(
diff --git a/benchmark/transformer/static/train.py b/benchmark/transformer/static/train.py
index 38818666f9bb5..fa33aac4de2e4 100644
--- a/benchmark/transformer/static/train.py
+++ b/benchmark/transformer/static/train.py
@@ -147,7 +147,7 @@ def do_train(args):
 
     if args.use_amp:
         optimizer.amp_init(places[0])
-    
+
     # the best cross-entropy value with label smoothing
     loss_normalizer = -(
         (1. - args.label_smooth_eps) * np.log(
@@ -181,6 +181,9 @@ def do_train(args):
                                    'lbl_word': data[i][2],
                                } for i in range(trainer_count)],
                                fetch_list=[sum_cost.name, token_num.name])
+                train_batch_cost = time.time() - batch_start
+                batch_ips_avg.record(train_batch_cost,
+                                     np.asarray(outs[1]).sum())
             else:
                 outs = exe.run(compiled_train_program,
                                feed=[{
@@ -189,12 +192,13 @@ def do_train(args):
                                    'lbl_word': data[i][2],
                                } for i in range(trainer_count)],
                                fetch_list=[sum_cost.name, token_num.name])
+                train_batch_cost = time.time() - batch_start
+                batch_ips_avg.record(train_batch_cost,
+                                     np.asarray(outs[1]).sum() / trainer_count)
             scheduler.step()
 
-            train_batch_cost = time.time() - batch_start
             reader_cost_avg.record(train_reader_cost)
             batch_cost_avg.record(train_batch_cost)
-            batch_ips_avg.record(train_batch_cost, np.asarray(outs[1]).sum())
 
             if step_idx % args.print_step == 0:
                 sum_cost_val, token_num_val = np.array(outs[0]), np.array(outs[