diff --git a/python/examples/bert/benchmark_batch.py b/python/examples/bert/benchmark_batch.py
index 265521d484259b0e6ea2b182dbf61e2a5cf43b8d..9b8e301a62eb0eee161cd701555543d329c6ae83 100644
--- a/python/examples/bert/benchmark_batch.py
+++ b/python/examples/bert/benchmark_batch.py
@@ -57,7 +57,7 @@ def single_func(idx, resource):
                         os.getpid(),
                         int(round(b_start * 1000000)),
                         int(round(b_end * 1000000))))
-                result = client.predict(feed_batch=feed_batch, fetch=fetch)
+                result = client.predict(feed=feed_batch, fetch=fetch)
             else:
                 print("unsupport batch size {}".format(args.batch_size))
 
diff --git a/python/examples/criteo_ctr/benchmark_batch.py b/python/examples/criteo_ctr/benchmark_batch.py
index ab67507355d0eba187d47ec9577eb5a3eda5dc46..1e4348c99dc0d960b1818ea6f0eb1ae2f5bd2ccb 100644
--- a/python/examples/criteo_ctr/benchmark_batch.py
+++ b/python/examples/criteo_ctr/benchmark_batch.py
@@ -55,7 +55,7 @@ def single_func(idx, resource):
                     for i in range(1, 27):
                         feed_dict["sparse_{}".format(i - 1)] = data[0][i]
                     feed_batch.append(feed_dict)
-                result = client.predict(feed_batch=feed_batch, fetch=fetch)
+                result = client.predict(feed=feed_batch, fetch=fetch)
             else:
                 print("unsupport batch size {}".format(args.batch_size))
 
diff --git a/python/examples/criteo_ctr_with_cube/benchmark_batch.py b/python/examples/criteo_ctr_with_cube/benchmark_batch.py
index b4b15892375e830486afa320151fac619aab2ba7..df5c6b90badb36fd7e349555973ccbd7ea0a8b70 100755
--- a/python/examples/criteo_ctr_with_cube/benchmark_batch.py
+++ b/python/examples/criteo_ctr_with_cube/benchmark_batch.py
@@ -56,8 +56,7 @@ def single_func(idx, resource):
                         feed_dict["embedding_{}.tmp_0".format(i - 1)] = data[0][
                             i]
                     feed_batch.append(feed_dict)
-                result = client.batch_predict(
-                    feed_batch=feed_batch, fetch=fetch)
+                result = client.predict(feed=feed_batch, fetch=fetch)
             else:
                 print("unsupport batch size {}".format(args.batch_size))
 
diff --git a/python/examples/imagenet/benchmark_batch.py b/python/examples/imagenet/benchmark_batch.py
index 915544d4f6f9d4636b14f4be92ad75ba08013389..e531425770cbf9102b7ebd2f5b082c5c4aa14e71 100644
--- a/python/examples/imagenet/benchmark_batch.py
+++ b/python/examples/imagenet/benchmark_batch.py
@@ -50,7 +50,7 @@ def single_func(idx, resource):
                     img = reader.process_image(img_list[i])
                     img = img.reshape(-1)
                     feed_batch.append({"image": img})
-                result = client.predict(feed_batch=feed_batch, fetch=fetch)
+                result = client.predict(feed=feed_batch, fetch=fetch)
             else:
                 print("unsupport batch size {}".format(args.batch_size))
 
diff --git a/python/examples/imdb/benchmark_batch.py b/python/examples/imdb/benchmark_batch.py
index 3ac52ec5472ff97f5d273dc230494223f3a71907..d36704a7631e963fd51220aa3c3d9a350515ebfd 100644
--- a/python/examples/imdb/benchmark_batch.py
+++ b/python/examples/imdb/benchmark_batch.py
@@ -42,8 +42,7 @@ def single_func(idx, resource):
                 for bi in range(args.batch_size):
                     word_ids, label = imdb_dataset.get_words_and_label(line)
                     feed_batch.append({"words": word_ids})
-                result = client.predict(
-                    feed_batch=feed_batch, fetch=["prediction"])
+                result = client.predict(feed=feed_batch, fetch=["prediction"])
             else:
                 print("unsupport batch size {}".format(args.batch_size))
 
diff --git a/python/paddle_serving_server_gpu/serve.py b/python/paddle_serving_server_gpu/serve.py
index 2cce4d4b1615584fc02aba4d70e1928083ddde62..d09efbfc8e1512ecb75b063ad760ce66e1a3159e 100644
--- a/python/paddle_serving_server_gpu/serve.py
+++ b/python/paddle_serving_server_gpu/serve.py
@@ -64,14 +64,22 @@ def start_gpu_card_model(index, gpuid, args):  # pylint: disable=doc-string-miss
 def start_multi_card(args):  # pylint: disable=doc-string-missing
     gpus = ""
     if args.gpu_ids == "":
-        if "CUDA_VISIBLE_DEVICES" in os.environ:
-            gpus = os.environ["CUDA_VISIBLE_DEVICES"]
-        else:
-            gpus = []
+        gpus = []
     else:
         gpus = args.gpu_ids.split(",")
+        if "CUDA_VISIBLE_DEVICES" in os.environ:
+            env_gpus = os.environ["CUDA_VISIBLE_DEVICES"].split(",")
+            for ids in gpus:
+                if int(ids) >= len(env_gpus):
+                    print(
+                        " Max index of gpu_ids out of range, the number of CUDA_VISIBLE_DEVICES is {}.".
+                        format(len(env_gpus)))
+                    exit(-1)
+        else:
+            env_gpus = []
     if len(gpus) <= 0:
-        start_gpu_card_model(-1, 0, args)
+        print("gpu_ids not set, going to run cpu service.")
+        start_gpu_card_model(-1, -1, args)
     else:
         gpu_processes = []
         for i, gpu_id in enumerate(gpus):