diff --git a/python/paddle_serving_server_gpu/__init__.py b/python/paddle_serving_server_gpu/__init__.py
index 31ba66cff68ce1e9f00b5d533cef67d3f2b08f71..ffa8e6eb5b2f3d38d2d2a6525a39356dd014e5d1 100644
--- a/python/paddle_serving_server_gpu/__init__.py
+++ b/python/paddle_serving_server_gpu/__init__.py
@@ -22,6 +22,7 @@ import socket
 import paddle_serving_server_gpu as paddle_serving_server
 from version import serving_server_version
 from contextlib import closing
+import argparse
 
 
 def serve_args():
diff --git a/python/paddle_serving_server_gpu/web_serve.py b/python/paddle_serving_server_gpu/web_serve.py
index c270997e228a07bc24a9214b93e8fe3494bcc1c7..de7130afb90b38cc867edab8134c297ad326989d 100644
--- a/python/paddle_serving_server_gpu/web_serve.py
+++ b/python/paddle_serving_server_gpu/web_serve.py
@@ -27,10 +27,13 @@ if __name__ == "__main__":
     args = serve_args()
     web_service = WebService(name=args.name)
     web_service.load_model_config(args.model)
+    gpu_ids = []
     if args.gpu_ids == "":
-        gpu_ids = os.environ["CUDA_VISIBLE_DEVICES"]
-    gpus = [int(x) for x in gpu_ids.split(",")]
-    web_service.set_gpus(gpus)
+        if "CUDA_VISIBLE_DEVICES" in os.environ:
+            gpu_ids = os.environ["CUDA_VISIBLE_DEVICES"]
+    if len(gpu_ids) > 0:
+        gpus = [int(x) for x in gpu_ids.split(",")]
+        web_service.set_gpus(gpus)
     web_service.prepare_server(
         workdir=args.workdir, port=args.port, device=args.device)
     service.run_server()
diff --git a/python/paddle_serving_server_gpu/web_service.py b/python/paddle_serving_server_gpu/web_service.py
index c90a6319c409d75cebe3f1bf87b14a3277f569a7..6d2039040178c118f48a09563735d97cca3c10c7 100755
--- a/python/paddle_serving_server_gpu/web_service.py
+++ b/python/paddle_serving_server_gpu/web_service.py
@@ -112,7 +112,7 @@ class WebService(object):
                 feed=request.json, fetch=fetch, fetch_map=fetch_map)
             return fetch_map
 
-        app_instance.run(host="127.0.0.1",
+        app_instance.run(host="0.0.0.0",
                          port=self.port,
                          threaded=False,
                          processes=1)