diff --git a/fluid/language_model/train.py b/fluid/language_model/train.py
index ac8dcd2949cd4054c091c3f64ce04164e2dcbe9b..f3e7a7398bf13e14c74ce1d10d90b7bf34031698 100644
--- a/fluid/language_model/train.py
+++ b/fluid/language_model/train.py
@@ -164,10 +164,13 @@ def train(train_reader,
     print("finish training")
 
 
-def get_cards():
-    cards = os.environ.get('CUDA_VISIBLE_DEVICES')
-    num = len(cards.split(","))
-    return num
+def get_cards(enable_ce):
+    if enable_ce:
+        cards = os.environ.get('CUDA_VISIBLE_DEVICES')
+        num = len(cards.split(","))
+        return num
+    else:
+        return fluid.core.get_cuda_device_count()
 
 
 def train_net():
@@ -175,7 +178,7 @@ def train_net():
     batch_size = 20
     args = parse_args()
     vocab, train_reader, test_reader = utils.prepare_data(
-        batch_size=batch_size * get_cards(), buffer_size=1000, \
+        batch_size=batch_size * get_cards(args.enable_ce), buffer_size=1000, \
         word_freq_threshold=0, enable_ce = args.enable_ce)
     train(
         train_reader=train_reader,