diff --git a/BERT/dist_utils.py b/BERT/dist_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..94cf415d0da5bba9ab68c7a7503b55fe68c095c4
--- /dev/null
+++ b/BERT/dist_utils.py
@@ -0,0 +1,45 @@
+#  Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+#
+#Licensed under the Apache License, Version 2.0 (the "License");
+#you may not use this file except in compliance with the License.
+#You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+#Unless required by applicable law or agreed to in writing, software
+#distributed under the License is distributed on an "AS IS" BASIS,
+#WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#See the License for the specific language governing permissions and
+#limitations under the License.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+import os
+import paddle.fluid as fluid
+
+def nccl2_prepare(trainer_id, startup_prog, main_prog):
+   config = fluid.DistributeTranspilerConfig()
+   config.mode = "nccl2"
+   t = fluid.DistributeTranspiler(config=config)
+   t.transpile(trainer_id,
+      trainers=os.environ.get('PADDLE_TRAINER_ENDPOINTS'),
+      current_endpoint=os.environ.get('PADDLE_CURRENT_ENDPOINT'),
+      startup_program=startup_prog,
+      program=main_prog)
+
+def prepare_for_multi_process(exe, build_strategy, train_prog, startup_prog):
+   # prepare for multi-process
+   trainer_id = int(os.environ.get('PADDLE_TRAINER_ID', 0))
+   num_trainers = int(os.environ.get('PADDLE_TRAINERS_NUM', 1))
+   print("PADDLE_TRAINERS_NUM", num_trainers)
+   print("PADDLE_TRAINER_ID", trainer_id)
+   build_strategy.num_trainers =  num_trainers
+   build_strategy.trainer_id = trainer_id
+   # NOTE(zcd): use multi processes to train the model,
+   # and each process use one GPU card.
+   if num_trainers > 1:
+       nccl2_prepare(trainer_id,
+           startup_prog, train_prog)
+       # the startup_prog are run two times, but it doesn't matter.
+       exe.run(startup_prog) 
diff --git a/BERT/run_classifier.py b/BERT/run_classifier.py
index 948ec3bc9eaf60f5c9fbcaad5b2cb856f8261133..f2ca68c4728b9c297c1ba25024dba51ab6f7acc5 100644
--- a/BERT/run_classifier.py
+++ b/BERT/run_classifier.py
@@ -32,6 +32,8 @@ from model.classifier import create_model
 from optimization import optimization
 from utils.args import ArgumentGroup, print_arguments
 from utils.init import init_pretraining_params, init_checkpoint
+import dist_utils
+
 
 # yapf: disable
 parser = argparse.ArgumentParser(__doc__)
@@ -107,6 +109,21 @@ def evaluate(exe, test_program, test_pyreader, fetch_list, eval_phase):
           (eval_phase, np.sum(total_cost) / np.sum(total_num_seqs),
            np.sum(total_acc) / np.sum(total_num_seqs), time_end - time_begin))
 
+def get_device_num():
+    visible_device = os.getenv('CUDA_VISIBLE_DEVICES')
+    # NOTE(zcd): use multi processes to train the model,
+    # and each process use one GPU card.
+    num_trainers = int(os.environ.get('PADDLE_TRAINERS_NUM', 1))
+    if num_trainers > 1 : return 1
+    if visible_device:
+        device_num = len(visible_device.split(','))
+    else:
+        device_num = subprocess.check_output(['nvidia-smi','-L']).decode().count('\n')
+    return device_num
+
+def update_lr(args):
+    num_trainers = int(os.environ.get('PADDLE_TRAINERS_NUM', 1))
+    args.learning_rate = args.learning_rate / num_trainers
 
 def main(args):
     bert_config = BertConfig(args.bert_config_path)
@@ -114,12 +131,14 @@ def main(args):
 
     if args.use_cuda:
         place = fluid.CUDAPlace(int(os.getenv('FLAGS_selected_gpus', '0')))
-        dev_count = fluid.core.get_cuda_device_count()
+        dev_count = get_device_num() # fluid.core.get_cuda_device_count()
     else:
         place = fluid.CPUPlace()
         dev_count = int(os.environ.get('CPU_NUM', multiprocessing.cpu_count()))
     exe = fluid.Executor(place)
 
+    update_lr(args)
+
     task_name = args.task_name.lower()
     processors = {
         'xnli': reader.XnliProcessor,
@@ -250,7 +269,9 @@ def main(args):
         exec_strategy.num_threads = dev_count
         exec_strategy.num_iteration_per_drop_scope = args.num_iteration_per_drop_scope
         build_strategy = fluid.BuildStrategy()
-        
+
+        dist_utils.prepare_for_multi_process(exe, build_strategy, train_program, startup_prog)
+
         train_exe = fluid.ParallelExecutor(
             use_cuda=args.use_cuda,
             loss_name=loss.name,