diff --git a/paddle/gserver/layers/GruStepLayer.cpp b/paddle/gserver/layers/GruStepLayer.cpp
index ce692c490881c9330b8113185bbec98b8dfd63e0..4a1006aa941f396c233a0cecfc38228f1f9fafe1 100644
--- a/paddle/gserver/layers/GruStepLayer.cpp
+++ b/paddle/gserver/layers/GruStepLayer.cpp
@@ -68,8 +68,8 @@ bool GruStepLayer::init(const LayerMap& layerMap,
   if (!Layer::init(layerMap, parameterMap)) return false;
   CHECK_EQ(2U, inputLayers_.size());
 
-  CHECK_EQ(getSize() * getSize() * 3, parameters_[1]->getSize());
-  weight_.reset(new Weight(getSize(), getSize() * 3, parameters_[1]));
+  CHECK_EQ(getSize() * getSize() * 3, parameters_[0]->getSize());
+  weight_.reset(new Weight(getSize(), getSize() * 3, parameters_[0]));
 
   if (biasParameter_.get() != NULL) {
     CHECK_EQ(getSize() * 3, biasParameter_->getSize());
diff --git a/paddle/gserver/tests/test_LayerGrad.cpp b/paddle/gserver/tests/test_LayerGrad.cpp
index 813a99d8cdae2e68473280ffc6e2370c02c9ec6c..8c8e876bd64fb97e11bc04c26ec45358f3f808a1 100644
--- a/paddle/gserver/tests/test_LayerGrad.cpp
+++ b/paddle/gserver/tests/test_LayerGrad.cpp
@@ -1404,9 +1404,9 @@ TEST(Layer, GruStepLayer) {
   config.biasSize = 12;
 
   config.inputDefs.push_back(
-      {INPUT_DATA, "layer_0", /* dim= */ 12, /* paraSize= */ 0});
+      {INPUT_DATA, "layer_0", /* dim= */ 12, /* paraSize= */ 48});
   config.inputDefs.push_back(
-      {INPUT_DATA, "layer_1", /* dim= */ 4, /* paraSize= */ 48});
+      {INPUT_DATA, "layer_1", /* dim= */ 4, /* paraSize= */ 0});
   config.layerConfig.add_inputs();
   config.layerConfig.add_inputs();
 
diff --git a/python/paddle/trainer/config_parser.py b/python/paddle/trainer/config_parser.py
index 4fbf076ae98c4d982329de1c1af7232c76ef8912..6701eced60d068312a1a866a6312002f9f5207f7 100644
--- a/python/paddle/trainer/config_parser.py
+++ b/python/paddle/trainer/config_parser.py
@@ -2996,7 +2996,7 @@ class GruStepLayer(LayerBase):
         config_assert(input_layer1.size == size,
                       'input_layer1.size != layer.size')
         self.config.active_gate_type = active_gate_type
-        self.create_input_parameter(1, size * size * 3, [size, size * 3])
+        self.create_input_parameter(0, size * size * 3, [size, size * 3])
         self.create_bias_parameter(bias, size * 3)
 
 
diff --git a/python/paddle/trainer_config_helpers/layers.py b/python/paddle/trainer_config_helpers/layers.py
index f0b5d7c3b4f26a723d7431eb1bfaa3a6eaec944f..14035ce3fd8aabe6a6800fa1be95193f053c1bba 100644
--- a/python/paddle/trainer_config_helpers/layers.py
+++ b/python/paddle/trainer_config_helpers/layers.py
@@ -2706,6 +2706,8 @@ def gru_step_layer(input,
     :param name:
     :param gate_act:
     :param bias_attr:
+    :param param_attr: the parameter_attribute for transforming the output_mem 
+                       from previous step. 
     :param layer_attr:
     :return: LayerOutput object.
     :rtype: LayerOutput
@@ -2716,7 +2718,12 @@ def gru_step_layer(input,
     Layer(
         name=name,
         type=LayerType.GRU_STEP_LAYER,
-        inputs=[input.name, Input(output_mem.name, **param_attr.attr)],
+        # The parameter here is for transforming the output_mem. The input has 
+        # already been transformed outside this module so it does not need 
+        # parameter associated with it. 
+        # The parameter here is instead grouped with input is due to 
+        # backward model compatibility.
+        inputs=[Input(input.name, **param_attr.attr), output_mem.name],
         bias=ParamAttr.to_bias(bias_attr),
         size=size,
         active_type=act.name,
diff --git a/python/paddle/trainer_config_helpers/tests/configs/protostr/shared_gru.protostr b/python/paddle/trainer_config_helpers/tests/configs/protostr/shared_gru.protostr
index c0868713ebbbc16c6cd3e557cb2af039cea70e0b..b6905824f0cb090375a38ff67e39fc626df0b2f6 100644
--- a/python/paddle/trainer_config_helpers/tests/configs/protostr/shared_gru.protostr
+++ b/python/paddle/trainer_config_helpers/tests/configs/protostr/shared_gru.protostr
@@ -51,10 +51,10 @@ layers {
   active_type: "tanh"
   inputs {
     input_layer_name: "__simple_gru_0___transform@__simple_gru_0___recurrent_group"
+    input_parameter_name: "gru_param"
   }
   inputs {
     input_layer_name: "__simple_gru_0__+delay1@__simple_gru_0___recurrent_group"
-    input_parameter_name: "gru_param"
   }
   bias_parameter_name: "gru_bias"
   active_gate_type: "sigmoid"
@@ -105,10 +105,10 @@ layers {
   active_type: "tanh"
   inputs {
     input_layer_name: "__simple_gru_1___transform@__simple_gru_1___recurrent_group"
+    input_parameter_name: "gru_param"
   }
   inputs {
     input_layer_name: "__simple_gru_1__+delay1@__simple_gru_1___recurrent_group"
-    input_parameter_name: "gru_param"
   }
   bias_parameter_name: "gru_bias"
   active_gate_type: "sigmoid"
diff --git a/python/paddle/trainer_config_helpers/tests/configs/protostr/test_rnn_group.protostr b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_rnn_group.protostr
index c1d39f77295892afc5ae12ec3f7c7c4ab6d5c0c2..3e9d28416ed5066461e960f0a9f085e057c28346 100644
--- a/python/paddle/trainer_config_helpers/tests/configs/protostr/test_rnn_group.protostr
+++ b/python/paddle/trainer_config_helpers/tests/configs/protostr/test_rnn_group.protostr
@@ -307,10 +307,10 @@ layers {
   active_type: "tanh"
   inputs {
     input_layer_name: "__mixed_1__@__gru_group_0___recurrent_group"
+    input_parameter_name: "___gru_group_0__@__gru_group_0___recurrent_group.w0"
   }
   inputs {
     input_layer_name: "__gru_group_0__+delay1@__gru_group_0___recurrent_group"
-    input_parameter_name: "___gru_group_0__@__gru_group_0___recurrent_group.w1"
   }
   bias_parameter_name: "___gru_group_0__@__gru_group_0___recurrent_group.wbias"
   active_gate_type: "sigmoid"
@@ -462,7 +462,7 @@ parameters {
   initial_smart: false
 }
 parameters {
-  name: "___gru_group_0__@__gru_group_0___recurrent_group.w1"
+  name: "___gru_group_0__@__gru_group_0___recurrent_group.w0"
   size: 30000
   initial_mean: 0.0
   initial_std: 0.1