diff --git a/python/paddle/v2/fluid/layers/control_flow.py b/python/paddle/v2/fluid/layers/control_flow.py
index 4b363ecbe78af82733fe1f80e44118a0dfda1f11..bef9602bb7cb94d45e663c7780a2b349c37a90b2 100644
--- a/python/paddle/v2/fluid/layers/control_flow.py
+++ b/python/paddle/v2/fluid/layers/control_flow.py
@@ -1220,7 +1220,8 @@ class DynamicRNN(object):
         self.lod_rank_table = None
         self.max_seq_len = None
         self.step_idx = None
-        self.zero_idx = fill_constant(shape=[1], value=0, dtype='int64')
+        self.zero_idx = fill_constant(
+            shape=[1], value=0, dtype='int64', force_cpu=True)
         self.mem_dict = dict()
         self.output_array = []
         self.outputs = []
@@ -1275,7 +1276,8 @@ class DynamicRNN(object):
     def block(self):
         if self.status != DynamicRNN.BEFORE_RNN:
             raise ValueError("rnn.block() can only be invoke once")
-        self.step_idx = fill_constant(shape=[1], dtype='int64', value=0)
+        self.step_idx = fill_constant(
+            shape=[1], dtype='int64', value=0, force_cpu=True)
         self.step_idx.stop_gradient = False
         self.status = DynamicRNN.IN_RNN
         with self.while_op.block():
diff --git a/python/paddle/v2/fluid/layers/tensor.py b/python/paddle/v2/fluid/layers/tensor.py
index 2608a8d1151fafa2da0cf7b605c4fa1210068057..2217c56b62af958709bb180bb360c938d1f0b312 100644
--- a/python/paddle/v2/fluid/layers/tensor.py
+++ b/python/paddle/v2/fluid/layers/tensor.py
@@ -180,7 +180,7 @@ def assign(input, output):
     return output
 
 
-def fill_constant(shape, dtype, value, out=None):
+def fill_constant(shape, dtype, value, force_cpu=False, out=None):
     """
     **fill_constant**
 
@@ -211,9 +211,12 @@ def fill_constant(shape, dtype, value, out=None):
         type='fill_constant',
         inputs={},
         outputs={'Out': [out]},
-        attrs={'shape': shape,
-               'dtype': out.dtype,
-               'value': float(value)})
+        attrs={
+            'shape': shape,
+            'dtype': out.dtype,
+            'value': float(value),
+            'force_cpu': force_cpu
+        })
     out.stop_gradient = True
     return out