diff --git a/python/paddle/hapi/model.py b/python/paddle/hapi/model.py
index 99e8acd2b0b934d2fd666472a8a6fa79c29277f6..137ca186d7946a426b263b6b902e101be4744135 100644
--- a/python/paddle/hapi/model.py
+++ b/python/paddle/hapi/model.py
@@ -621,6 +621,7 @@ class DynamicGraphAdapter(object):
 
         self._input_info = None
         if self._nranks > 1:
+            dist.init_parallel_env()
             stradegy = fluid.dygraph.parallel.ParallelStrategy()
             stradegy.nranks = ParallelEnv().nranks
             stradegy.local_rank = ParallelEnv().local_rank
@@ -888,7 +889,6 @@ class Model(object):
 
         # init backend
         if fluid.in_dygraph_mode():
-            dist.init_parallel_env()
             self._adapter = DynamicGraphAdapter(self)
         else:
             self._adapter = StaticGraphAdapter(self)
@@ -943,6 +943,7 @@ class Model(object):
             self._update_inputs()
         return loss
 
+    @paddle.no_grad()
     def eval_batch(self, inputs, labels=None):
         """
         Run one evaluating step on a batch of data.
@@ -994,6 +995,7 @@ class Model(object):
             self._update_inputs()
         return loss
 
+    @paddle.no_grad()
     def predict_batch(self, inputs):
         """
         Run one predicting step on a batch of data.
diff --git a/python/paddle/tests/dist_hapi_mnist_dynamic.py b/python/paddle/tests/dist_hapi_mnist_dynamic.py
index 46d02789402b22263cfbd8cbdfeb6d66a5de900d..eab34a6dafbc354a24aa51e93a9fec9efc3b3cee 100644
--- a/python/paddle/tests/dist_hapi_mnist_dynamic.py
+++ b/python/paddle/tests/dist_hapi_mnist_dynamic.py
@@ -61,7 +61,6 @@ class TestDistTraning(unittest.TestCase):
     def test_static_multiple_gpus(self):
         device = set_device('gpu')
 
-        fluid.enable_dygraph(device)
         im_shape = (-1, 1, 28, 28)
         batch_size = 128