diff --git a/paddle/fluid/framework/op_desc.cc b/paddle/fluid/framework/op_desc.cc
index 121e00b1a3d761f570138092a3e76ae2b722d28a..c293cf92b4f3d530109c76850df184af9cad7399 100644
--- a/paddle/fluid/framework/op_desc.cc
+++ b/paddle/fluid/framework/op_desc.cc
@@ -515,20 +515,14 @@ void OpDesc::InferShape(const BlockDesc &block) const {
 }
 
 void OpDesc::InferVarType(BlockDesc *block) const {
+  // There are a few places that var type can be set.
+  // When VarDesc is created, default set to LOD_TENSOR.
+  // When output variable is created, default is defaut set to LOD_TENSOR.
+  // We limit here to be the only place that operator defines its customized
+  // var type inference. Hence, we don't do any "default" setting here.
   auto &info = OpInfoMap::Instance().Get(this->Type());
   if (info.infer_var_type_) {
     info.infer_var_type_(*this, block);
-  } else {
-    // all output type is LoDTensor by default
-    VLOG(10) << this->Type()
-             << " has not registered InferVarType. Set output variables to "
-                "LOD_TENSOR";
-    for (auto &out_pair : this->outputs_) {
-      for (auto &out_var_name : out_pair.second) {
-        block->FindRecursiveOrCreateVar(out_var_name)
-            .SetType(proto::VarType::LOD_TENSOR);
-      }
-    }
   }
 }
 
diff --git a/python/paddle/fluid/layer_helper.py b/python/paddle/fluid/layer_helper.py
index bd9727b6ac0208b199091db00bd0fd5fae74d53b..dc317de9abbd06f4021e64b87ea88ba6af8809c9 100644
--- a/python/paddle/fluid/layer_helper.py
+++ b/python/paddle/fluid/layer_helper.py
@@ -324,10 +324,19 @@ class LayerHelper(object):
             raise ValueError("no Parameter name %s found" % name)
         return param
 
-    def create_tmp_variable(self, dtype, stop_gradient=False):
+    def create_variable_for_type_inference(self, dtype, stop_gradient=False):
+        """Create a temporary variable that should be type inferred layer.
+
+        Note:
+            The default type will be set to LOD_TENSOR. However, when
+            the var is used as operator output, its type will be updated
+            based on operator's `VarTypeInference` implementation in
+            infer_var_type.
+        """
         return self.main_program.current_block().create_var(
             name=unique_name.generate(".".join([self.name, 'tmp'])),
             dtype=dtype,
+            type=core.VarDesc.VarType.LOD_TENSOR,
             persistable=False,
             stop_gradient=stop_gradient)
 
@@ -388,7 +397,7 @@ class LayerHelper(object):
 
         b = self.create_parameter(
             attr=bias_attr, shape=size, dtype=input_var.dtype, is_bias=True)
-        tmp = self.create_tmp_variable(dtype=input_var.dtype)
+        tmp = self.create_variable_for_type_inference(dtype=input_var.dtype)
         self.append_op(
             type='elementwise_add',
             inputs={'X': [input_var],
@@ -414,7 +423,7 @@ class LayerHelper(object):
         tmp = input_var
         # NOTE(dzhwinter): some activation support inplace compution.
         if not core.IsInplace(act_type):
-            tmp = self.create_tmp_variable(dtype=input_var.dtype)
+            tmp = self.create_variable_for_type_inference(dtype=input_var.dtype)
         self.append_op(
             type=act_type,
             inputs={"X": [input_var]},
diff --git a/python/paddle/fluid/layers/control_flow.py b/python/paddle/fluid/layers/control_flow.py
index 4af97e8632a47fbd981362dc8249a3f6b7269ecd..459be4339b01228ba1b8f18e17e472ed72928511 100644
--- a/python/paddle/fluid/layers/control_flow.py
+++ b/python/paddle/fluid/layers/control_flow.py
@@ -80,8 +80,8 @@ def split_lod_tensor(input, mask, level=0):
 
     """
     helper = LayerHelper('split_lod_tensor', **locals())
-    out_true = helper.create_tmp_variable(dtype=input.dtype)
-    out_false = helper.create_tmp_variable(dtype=input.dtype)
+    out_true = helper.create_variable_for_type_inference(dtype=input.dtype)
+    out_false = helper.create_variable_for_type_inference(dtype=input.dtype)
     helper.append_op(
         type='split_lod_tensor',
         inputs={
@@ -131,7 +131,7 @@ def merge_lod_tensor(in_true, in_false, x, mask, level=0):
                 in_true=out_true, in_false=out_false, mask=y, x=x, level=level)
     """
     helper = LayerHelper('merge_lod_tensor', **locals())
-    out = helper.create_tmp_variable(dtype=in_true.dtype)
+    out = helper.create_variable_for_type_inference(dtype=in_true.dtype)
     helper.append_op(
         type='merge_lod_tensor',
         inputs={'X': x,
@@ -524,7 +524,7 @@ class StaticRNN(object):
         if not isinstance(o, Variable):
             raise TypeError("step output takes a Variable")
 
-        tmp_o = self.helper.create_tmp_variable(dtype=o.dtype)
+        tmp_o = self.helper.create_variable_for_type_inference(dtype=o.dtype)
         self.helper.append_op(
             type='rnn_memory_helper',
             inputs={'X': [o]},
@@ -606,7 +606,8 @@ class StaticRNN(object):
             pre_memories.append(mem.pre_mem.name)
             mem_var = rnn_block.var(mem.mem.name)
             assert isinstance(mem_var, Variable)
-            new_mem = self.helper.create_tmp_variable(dtype=mem_var.dtype)
+            new_mem = self.helper.create_variable_for_type_inference(
+                dtype=mem_var.dtype)
 
             rnn_block.append_op(
                 type='rnn_memory_helper',
@@ -813,7 +814,7 @@ def max_sequence_len(rank_table):
         ${out_comment}.
     """
     helper = LayerHelper("max_seqence_len", **locals())
-    res = helper.create_tmp_variable(dtype="int64")
+    res = helper.create_variable_for_type_inference(dtype="int64")
     helper.append_op(
         type="max_sequence_len",
         inputs={"RankTable": rank_table},
@@ -884,7 +885,7 @@ def array_to_lod_tensor(x, table):
           lod_tensor = fluid.layers.array_to_lod_tensor(array, table)
     """
     helper = LayerHelper("array_to_lod_tensor", **locals())
-    tmp = helper.create_tmp_variable(dtype=x.dtype)
+    tmp = helper.create_variable_for_type_inference(dtype=x.dtype)
     helper.append_op(
         type="array_to_lod_tensor",
         inputs={'X': x,
@@ -915,7 +916,7 @@ def increment(x, value=1.0, in_place=True):
     """
     helper = LayerHelper("increment", **locals())
     if not in_place:
-        out = helper.create_tmp_variable(dtype=x.dtype)
+        out = helper.create_variable_for_type_inference(dtype=x.dtype)
     else:
         out = x
     helper.append_op(
@@ -1012,7 +1013,7 @@ def less_than(x, y, force_cpu=None, cond=None, **ignored):
     """
     helper = LayerHelper("less_than", **locals())
     if cond is None:
-        cond = helper.create_tmp_variable(dtype='bool')
+        cond = helper.create_variable_for_type_inference(dtype='bool')
         cond.stop_gradient = True
 
     attrs = dict()
@@ -1051,7 +1052,7 @@ def equal(x, y, cond=None, **ignored):
     """
     helper = LayerHelper("equal", **locals())
     if cond is None:
-        cond = helper.create_tmp_variable(dtype='bool')
+        cond = helper.create_variable_for_type_inference(dtype='bool')
         cond.stop_gradient = True
 
     helper.append_op(
@@ -1098,7 +1099,7 @@ def array_read(array, i):
             array,
             Variable) or array.type != core.VarDesc.VarType.LOD_TENSOR_ARRAY:
         raise TypeError("array should be tensor array vairable")
-    out = helper.create_tmp_variable(dtype=array.dtype)
+    out = helper.create_variable_for_type_inference(dtype=array.dtype)
     helper.append_op(
         type='read_from_array',
         inputs={'X': [array],
@@ -1133,7 +1134,7 @@ def shrink_memory(x, i, table):
         usage.
     """
     helper = LayerHelper('shrink_memory', **locals())
-    out = helper.create_tmp_variable(dtype=x.dtype)
+    out = helper.create_variable_for_type_inference(dtype=x.dtype)
     helper.append_op(
         type='shrink_rnn_memory',
         inputs={'X': [x],
@@ -1170,7 +1171,7 @@ def array_length(array):
 
     """
     helper = LayerHelper('array_length', **locals())
-    tmp = helper.create_tmp_variable(dtype='int64')
+    tmp = helper.create_variable_for_type_inference(dtype='int64')
     tmp.stop_gradient = True
     helper.append_op(
         type='lod_array_length', inputs={'X': [array]}, outputs={'Out': [tmp]})
@@ -1590,7 +1591,7 @@ class DynamicRNN(object):
         self.mem_dict = dict()
         self.output_array = []
         self.outputs = []
-        self.cond = self.helper.create_tmp_variable(dtype='bool')
+        self.cond = self.helper.create_variable_for_type_inference(dtype='bool')
         self.cond.stop_gradient = False
         self.while_op = While(self.cond)
         self.input_array = []
@@ -1924,7 +1925,7 @@ def reorder_lod_tensor_by_rank(x, rank_table):
     helper.is_instance('x', Variable)
     helper.is_instance('rank_table', Variable)
 
-    out = helper.create_tmp_variable(dtype=x.dtype)
+    out = helper.create_variable_for_type_inference(dtype=x.dtype)
     helper.append_op(
         type='reorder_lod_tensor_by_rank',
         inputs={'X': [x],
@@ -1958,7 +1959,7 @@ def is_empty(x, cond=None, **ignored):
     """
     helper = LayerHelper("is_empty", **locals())
     if cond is None:
-        cond = helper.create_tmp_variable(dtype='bool')
+        cond = helper.create_variable_for_type_inference(dtype='bool')
         cond.stop_gradient = True
     elif not isinstance(cond, Variable):
         raise TypeError("cond takes a variable")
diff --git a/python/paddle/fluid/layers/detection.py b/python/paddle/fluid/layers/detection.py
index 1cfcbbb9c1614f21848e62cce79befc673e1739c..b94b59631a3d8999f569acf1027c71d3019f5c56 100644
--- a/python/paddle/fluid/layers/detection.py
+++ b/python/paddle/fluid/layers/detection.py
@@ -147,10 +147,11 @@ def rpn_target_assign(bbox_pred,
 
     helper = LayerHelper('rpn_target_assign', **locals())
     # Assign target label to anchors
-    loc_index = helper.create_tmp_variable(dtype='int32')
-    score_index = helper.create_tmp_variable(dtype='int32')
-    target_label = helper.create_tmp_variable(dtype='int32')
-    target_bbox = helper.create_tmp_variable(dtype=anchor_box.dtype)
+    loc_index = helper.create_variable_for_type_inference(dtype='int32')
+    score_index = helper.create_variable_for_type_inference(dtype='int32')
+    target_label = helper.create_variable_for_type_inference(dtype='int32')
+    target_bbox = helper.create_variable_for_type_inference(
+        dtype=anchor_box.dtype)
     helper.append_op(
         type="rpn_target_assign",
         inputs={
@@ -282,7 +283,8 @@ def detection_output(loc,
     scores = nn.reshape(x=scores, shape=compile_shape, actual_shape=run_shape)
     scores = nn.transpose(scores, perm=[0, 2, 1])
     scores.stop_gradient = True
-    nmsed_outs = helper.create_tmp_variable(dtype=decoded_box.dtype)
+    nmsed_outs = helper.create_variable_for_type_inference(
+        dtype=decoded_box.dtype)
     helper.append_op(
         type="multiclass_nms",
         inputs={'Scores': scores,
@@ -314,7 +316,7 @@ def iou_similarity(x, y, name=None):
     """
     helper = LayerHelper("iou_similarity", **locals())
     if name is None:
-        out = helper.create_tmp_variable(dtype=x.dtype)
+        out = helper.create_variable_for_type_inference(dtype=x.dtype)
     else:
         out = helper.create_variable(
             name=name, dtype=x.dtype, persistable=False)
@@ -351,7 +353,8 @@ def box_coder(prior_box,
     helper = LayerHelper("box_coder", **locals())
 
     if name is None:
-        output_box = helper.create_tmp_variable(dtype=prior_box.dtype)
+        output_box = helper.create_variable_for_type_inference(
+            dtype=prior_box.dtype)
     else:
         output_box = helper.create_variable(
             name=name, dtype=prior_box.dtype, persistable=False)
@@ -382,7 +385,7 @@ def polygon_box_transform(input, name=None):
     """
     helper = LayerHelper("polygon_box_transform", **locals())
     if name is None:
-        output = helper.create_tmp_variable(dtype=input.dtype)
+        output = helper.create_variable_for_type_inference(dtype=input.dtype)
     else:
         output = helper.create_variable(
             name=name, dtype=prior_box.input, persistable=False)
@@ -450,7 +453,7 @@ def detection_map(detect_res,
     helper = LayerHelper("detection_map", **locals())
 
     def __create_var(type):
-        return helper.create_tmp_variable(dtype=type)
+        return helper.create_variable_for_type_inference(dtype=type)
 
     map_out = __create_var('float32')
     accum_pos_count_out = out_states[0] if out_states else __create_var('int32')
@@ -557,8 +560,9 @@ def bipartite_match(dist_matrix,
         >>> matched_indices, matched_dist = fluid.layers.bipartite_match(iou)
     """
     helper = LayerHelper('bipartite_match', **locals())
-    match_indices = helper.create_tmp_variable(dtype='int32')
-    match_distance = helper.create_tmp_variable(dtype=dist_matrix.dtype)
+    match_indices = helper.create_variable_for_type_inference(dtype='int32')
+    match_distance = helper.create_variable_for_type_inference(
+        dtype=dist_matrix.dtype)
     helper.append_op(
         type='bipartite_match',
         inputs={'DistMat': dist_matrix},
@@ -644,8 +648,8 @@ def target_assign(input,
                             gt, matched_indices, mismatch_value=0)
     """
     helper = LayerHelper('target_assign', **locals())
-    out = helper.create_tmp_variable(dtype=input.dtype)
-    out_weight = helper.create_tmp_variable(dtype='float32')
+    out = helper.create_variable_for_type_inference(dtype=input.dtype)
+    out_weight = helper.create_variable_for_type_inference(dtype='float32')
     helper.append_op(
         type='target_assign',
         inputs={
@@ -816,9 +820,10 @@ def ssd_loss(location,
     conf_loss = nn.reshape(
         x=conf_loss, shape=(num, num_prior), actual_shape=actual_shape)
     conf_loss.stop_gradient = True
-    neg_indices = helper.create_tmp_variable(dtype='int32')
+    neg_indices = helper.create_variable_for_type_inference(dtype='int32')
     dtype = matched_indices.dtype
-    updated_matched_indices = helper.create_tmp_variable(dtype=dtype)
+    updated_matched_indices = helper.create_variable_for_type_inference(
+        dtype=dtype)
     helper.append_op(
         type='mine_hard_examples',
         inputs={
@@ -998,8 +1003,8 @@ def prior_box(input,
             max_sizes = [max_sizes]
         attrs['max_sizes'] = max_sizes
 
-    box = helper.create_tmp_variable(dtype)
-    var = helper.create_tmp_variable(dtype)
+    box = helper.create_variable_for_type_inference(dtype)
+    var = helper.create_variable_for_type_inference(dtype)
     helper.append_op(
         type="prior_box",
         inputs={"Input": input,
@@ -1337,8 +1342,8 @@ def anchor_generator(input,
         'offset': offset
     }
 
-    anchor = helper.create_tmp_variable(dtype)
-    var = helper.create_tmp_variable(dtype)
+    anchor = helper.create_variable_for_type_inference(dtype)
+    var = helper.create_variable_for_type_inference(dtype)
     helper.append_op(
         type="anchor_generator",
         inputs={"Input": input},
@@ -1384,7 +1389,7 @@ def roi_perspective_transform(input,
     """
     helper = LayerHelper('roi_perspective_transform', **locals())
     dtype = helper.input_dtype()
-    out = helper.create_tmp_variable(dtype)
+    out = helper.create_variable_for_type_inference(dtype)
     helper.append_op(
         type="roi_perspective_transform",
         inputs={"X": input,
@@ -1418,11 +1423,15 @@ def generate_proposal_labels(rpn_rois,
 
     helper = LayerHelper('generate_proposal_labels', **locals())
 
-    rois = helper.create_tmp_variable(dtype=rpn_rois.dtype)
-    labels_int32 = helper.create_tmp_variable(dtype=gt_classes.dtype)
-    bbox_targets = helper.create_tmp_variable(dtype=rpn_rois.dtype)
-    bbox_inside_weights = helper.create_tmp_variable(dtype=rpn_rois.dtype)
-    bbox_outside_weights = helper.create_tmp_variable(dtype=rpn_rois.dtype)
+    rois = helper.create_variable_for_type_inference(dtype=rpn_rois.dtype)
+    labels_int32 = helper.create_variable_for_type_inference(
+        dtype=gt_classes.dtype)
+    bbox_targets = helper.create_variable_for_type_inference(
+        dtype=rpn_rois.dtype)
+    bbox_inside_weights = helper.create_variable_for_type_inference(
+        dtype=rpn_rois.dtype)
+    bbox_outside_weights = helper.create_variable_for_type_inference(
+        dtype=rpn_rois.dtype)
 
     helper.append_op(
         type="generate_proposal_labels",
@@ -1504,8 +1513,10 @@ def generate_proposals(scores,
     """
     helper = LayerHelper('generate_proposals', **locals())
 
-    rpn_rois = helper.create_tmp_variable(dtype=bbox_deltas.dtype)
-    rpn_roi_probs = helper.create_tmp_variable(dtype=scores.dtype)
+    rpn_rois = helper.create_variable_for_type_inference(
+        dtype=bbox_deltas.dtype)
+    rpn_roi_probs = helper.create_variable_for_type_inference(
+        dtype=scores.dtype)
     helper.append_op(
         type="generate_proposals",
         inputs={
diff --git a/python/paddle/fluid/layers/io.py b/python/paddle/fluid/layers/io.py
index dcd5a064a85cbfeedf41b7dc122cacba987e6570..95e13669ad927cdda56153c7fe58325f4a3ec474 100644
--- a/python/paddle/fluid/layers/io.py
+++ b/python/paddle/fluid/layers/io.py
@@ -954,7 +954,7 @@ def read_file(reader):
     """
     helper = LayerHelper('read_file')
     out = [
-        helper.create_tmp_variable(
+        helper.create_variable_for_type_inference(
             stop_gradient=True, dtype='float32')
         for _ in range(len(reader.desc.shapes()))
     ]
diff --git a/python/paddle/fluid/layers/layer_function_generator.py b/python/paddle/fluid/layers/layer_function_generator.py
index 8c11921d9bde0920f33368837302d39f36f45556..eea0a362a0c31083f304a2167d0fdadfb30fb640 100644
--- a/python/paddle/fluid/layers/layer_function_generator.py
+++ b/python/paddle/fluid/layers/layer_function_generator.py
@@ -202,10 +202,12 @@ def generate_layer_fn(op_type):
             out_var = out[0] if (isinstance(out, list) or
                                  isinstance(out, tuple)) else out
         else:
-            out_var = helper.create_tmp_variable(dtype=dtype)
+            out_var = helper.create_variable_for_type_inference(dtype=dtype)
         outputs[o_name] = [out_var]
         for name in intermediate_output_names:
-            outputs[name] = [helper.create_tmp_variable(dtype=dtype)]
+            outputs[name] = [
+                helper.create_variable_for_type_inference(dtype=dtype)
+            ]
         helper.append_op(
             type=op_type, inputs=inputs, outputs=outputs, attrs=kwargs)
         return helper.append_activation(out_var)
@@ -229,7 +231,7 @@ def generate_layer_fn_noattr(op_type):
 
     def func(x, name=None):
         helper = LayerHelper(op_type, **locals())
-        output = helper.create_tmp_variable(dtype=x.dtype)
+        output = helper.create_variable_for_type_inference(dtype=x.dtype)
         helper.append_op(type=op_type, inputs={"X": x}, outputs={"Out": output})
         return output
 
diff --git a/python/paddle/fluid/layers/metric_op.py b/python/paddle/fluid/layers/metric_op.py
index a3064b565d096f7feda18379c66ffc8bf2f4a55c..b2d2c93ead80d781d0a55ca541a1b0bb4232ad81 100644
--- a/python/paddle/fluid/layers/metric_op.py
+++ b/python/paddle/fluid/layers/metric_op.py
@@ -58,11 +58,11 @@ def accuracy(input, label, k=1, correct=None, total=None):
     """
     helper = LayerHelper("accuracy", **locals())
     topk_out, topk_indices = nn.topk(input, k=k)
-    acc_out = helper.create_tmp_variable(dtype="float32")
+    acc_out = helper.create_variable_for_type_inference(dtype="float32")
     if correct is None:
-        correct = helper.create_tmp_variable(dtype="int64")
+        correct = helper.create_variable_for_type_inference(dtype="int64")
     if total is None:
-        total = helper.create_tmp_variable(dtype="int64")
+        total = helper.create_variable_for_type_inference(dtype="int64")
     helper.append_op(
         type="accuracy",
         inputs={
@@ -124,8 +124,8 @@ def auc(input,
             auc_out=fluid.layers.auc(input=prediction, label=label)
     """
     helper = LayerHelper("auc", **locals())
-    auc_out = helper.create_tmp_variable(dtype="float64")
-    batch_auc_out = helper.create_tmp_variable(dtype="float64")
+    auc_out = helper.create_variable_for_type_inference(dtype="float64")
+    batch_auc_out = helper.create_variable_for_type_inference(dtype="float64")
     # make tp, tn, fp, fn persistable, so that can accumulate all batches.
 
     # for batch auc
diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py
index fb4ff9182f127daca179a599f4335a6b26622528..fc116445beeb4f68fe78c6f4c76e0ef688808e24 100644
--- a/python/paddle/fluid/layers/nn.py
+++ b/python/paddle/fluid/layers/nn.py
@@ -242,7 +242,7 @@ def fc(input,
 
         w = helper.create_parameter(
             attr=param_attr, shape=param_shape, dtype=dtype, is_bias=False)
-        tmp = helper.create_tmp_variable(dtype)
+        tmp = helper.create_variable_for_type_inference(dtype)
         helper.append_op(
             type="mul",
             inputs={"X": input_var,
@@ -255,7 +255,7 @@ def fc(input,
     if len(mul_results) == 1:
         pre_bias = mul_results[0]
     else:
-        pre_bias = helper.create_tmp_variable(dtype)
+        pre_bias = helper.create_variable_for_type_inference(dtype)
         helper.append_op(
             type="sum",
             inputs={"X": mul_results},
@@ -314,7 +314,7 @@ def embedding(input,
     helper = LayerHelper('embedding', **locals())
     w = helper.create_parameter(
         attr=helper.param_attr, shape=size, dtype=dtype, is_bias=False)
-    tmp = helper.create_tmp_variable(dtype)
+    tmp = helper.create_variable_for_type_inference(dtype)
     padding_idx = -1 if padding_idx is None else padding_idx if padding_idx >= 0 else (
         size[0] + padding_idx)
     helper.append_op(
@@ -418,10 +418,10 @@ def dynamic_lstm(input,
     bias = helper.create_parameter(
         attr=helper.bias_attr, shape=bias_size, dtype=dtype, is_bias=True)
 
-    hidden = helper.create_tmp_variable(dtype)
-    cell = helper.create_tmp_variable(dtype)
-    batch_gate = helper.create_tmp_variable(dtype)
-    batch_cell_pre_act = helper.create_tmp_variable(dtype)
+    hidden = helper.create_variable_for_type_inference(dtype)
+    cell = helper.create_variable_for_type_inference(dtype)
+    batch_gate = helper.create_variable_for_type_inference(dtype)
+    batch_cell_pre_act = helper.create_variable_for_type_inference(dtype)
     inputs = {'Input': input, 'Weight': weight, 'Bias': bias}
     batch_size = input.shape[0]
     if h_0:
@@ -621,12 +621,12 @@ def dynamic_lstmp(input,
     bias = helper.create_parameter(
         attr=helper.bias_attr, shape=bias_size, dtype=dtype, is_bias=True)
 
-    projection = helper.create_tmp_variable(dtype)
-    cell = helper.create_tmp_variable(dtype)
-    ordered_proj0 = helper.create_tmp_variable(dtype)
-    batch_hidden = helper.create_tmp_variable(dtype)
-    batch_gate = helper.create_tmp_variable(dtype)
-    batch_cell_pre_act = helper.create_tmp_variable(dtype)
+    projection = helper.create_variable_for_type_inference(dtype)
+    cell = helper.create_variable_for_type_inference(dtype)
+    ordered_proj0 = helper.create_variable_for_type_inference(dtype)
+    batch_hidden = helper.create_variable_for_type_inference(dtype)
+    batch_gate = helper.create_variable_for_type_inference(dtype)
+    batch_cell_pre_act = helper.create_variable_for_type_inference(dtype)
 
     helper.append_op(
         type='lstmp',
@@ -751,10 +751,10 @@ def dynamic_gru(input,
         ), 'The shape of h0 should be(batch_size, %d)' % size
         inputs['H0'] = h_0
 
-    hidden = helper.create_tmp_variable(dtype)
-    batch_gate = helper.create_tmp_variable(dtype)
-    batch_reset_hidden_prev = helper.create_tmp_variable(dtype)
-    batch_hidden = helper.create_tmp_variable(dtype)
+    hidden = helper.create_variable_for_type_inference(dtype)
+    batch_gate = helper.create_variable_for_type_inference(dtype)
+    batch_reset_hidden_prev = helper.create_variable_for_type_inference(dtype)
+    batch_hidden = helper.create_variable_for_type_inference(dtype)
 
     helper.append_op(
         type='gru',
@@ -844,9 +844,9 @@ def gru_unit(input,
     weight = helper.create_parameter(
         attr=helper.param_attr, shape=[size, 3 * size], dtype=dtype)
 
-    gate = helper.create_tmp_variable(dtype)
-    reset_hidden_pre = helper.create_tmp_variable(dtype)
-    updated_hidden = helper.create_tmp_variable(dtype)
+    gate = helper.create_variable_for_type_inference(dtype)
+    reset_hidden_pre = helper.create_variable_for_type_inference(dtype)
+    updated_hidden = helper.create_variable_for_type_inference(dtype)
     inputs = {'Input': input, 'HiddenPrev': hidden, 'Weight': weight}
     # create bias
     if helper.bias_attr:
@@ -896,10 +896,14 @@ def linear_chain_crf(input, label, param_attr=None):
         attr=helper.param_attr,
         shape=[size + 2, size],
         dtype=helper.input_dtype())
-    alpha = helper.create_tmp_variable(dtype=helper.input_dtype())
-    emission_exps = helper.create_tmp_variable(dtype=helper.input_dtype())
-    transition_exps = helper.create_tmp_variable(dtype=helper.input_dtype())
-    log_likelihood = helper.create_tmp_variable(dtype=helper.input_dtype())
+    alpha = helper.create_variable_for_type_inference(
+        dtype=helper.input_dtype())
+    emission_exps = helper.create_variable_for_type_inference(
+        dtype=helper.input_dtype())
+    transition_exps = helper.create_variable_for_type_inference(
+        dtype=helper.input_dtype())
+    log_likelihood = helper.create_variable_for_type_inference(
+        dtype=helper.input_dtype())
     helper.append_op(
         type='linear_chain_crf',
         inputs={"Emission": [input],
@@ -938,7 +942,8 @@ def crf_decoding(input, param_attr, label=None):
     """
     helper = LayerHelper('crf_decoding', **locals())
     transition = helper.get_parameter(param_attr.name)
-    viterbi_path = helper.create_tmp_variable(dtype=helper.input_dtype())
+    viterbi_path = helper.create_variable_for_type_inference(
+        dtype=helper.input_dtype())
     helper.append_op(
         type='crf_decoding',
         inputs={"Emission": [input],
@@ -962,9 +967,9 @@ def cos_sim(X, Y):
         Variable: the output of cosine(X, Y).
     """
     helper = LayerHelper('cos_sim', **locals())
-    out = helper.create_tmp_variable(dtype=X.dtype)
-    xnorm = helper.create_tmp_variable(dtype=X.dtype)
-    ynorm = helper.create_tmp_variable(dtype=X.dtype)
+    out = helper.create_variable_for_type_inference(dtype=X.dtype)
+    xnorm = helper.create_variable_for_type_inference(dtype=X.dtype)
+    ynorm = helper.create_variable_for_type_inference(dtype=X.dtype)
     helper.append_op(
         type='cos_sim',
         inputs={'X': [X],
@@ -1008,8 +1013,9 @@ def dropout(x, dropout_prob, is_test=False, seed=None, name=None):
     """
 
     helper = LayerHelper('dropout', **locals())
-    out = helper.create_tmp_variable(dtype=x.dtype)
-    mask = helper.create_tmp_variable(dtype=x.dtype, stop_gradient=True)
+    out = helper.create_variable_for_type_inference(dtype=x.dtype)
+    mask = helper.create_variable_for_type_inference(
+        dtype=x.dtype, stop_gradient=True)
 
     if (seed is None or seed == 0) and helper.main_program.random_seed != 0:
         seed = helper.main_program.random_seed
@@ -1094,7 +1100,7 @@ def cross_entropy(input, label, soft_label=False, ignore_index=-100):
           cost = fluid.layers.cross_entropy(input=predict, label=label)
     """
     helper = LayerHelper('cross_entropy', **locals())
-    out = helper.create_tmp_variable(dtype=input.dtype)
+    out = helper.create_variable_for_type_inference(dtype=input.dtype)
     helper.append_op(
         type='cross_entropy',
         inputs={'X': [input],
@@ -1141,14 +1147,14 @@ def square_error_cost(input, label):
 
     """
     helper = LayerHelper('square_error_cost', **locals())
-    minus_out = helper.create_tmp_variable(dtype=input.dtype)
+    minus_out = helper.create_variable_for_type_inference(dtype=input.dtype)
     helper.append_op(
         type='elementwise_sub',
         inputs={'X': [input],
                 'Y': [label]},
         outputs={'Out': [minus_out]})
 
-    square_out = helper.create_tmp_variable(dtype=input.dtype)
+    square_out = helper.create_variable_for_type_inference(dtype=input.dtype)
     helper.append_op(
         type='square', inputs={'X': [minus_out]},
         outputs={'Out': [square_out]})
@@ -1254,12 +1260,13 @@ def chunk_eval(input,
     helper = LayerHelper("chunk_eval", **locals())
 
     # prepare output
-    precision = helper.create_tmp_variable(dtype="float32")
-    recall = helper.create_tmp_variable(dtype="float32")
-    f1_score = helper.create_tmp_variable(dtype="float32")
-    num_infer_chunks = helper.create_tmp_variable(dtype="int64")
-    num_label_chunks = helper.create_tmp_variable(dtype="int64")
-    num_correct_chunks = helper.create_tmp_variable(dtype="int64")
+    precision = helper.create_variable_for_type_inference(dtype="float32")
+    recall = helper.create_variable_for_type_inference(dtype="float32")
+    f1_score = helper.create_variable_for_type_inference(dtype="float32")
+    num_infer_chunks = helper.create_variable_for_type_inference(dtype="int64")
+    num_label_chunks = helper.create_variable_for_type_inference(dtype="int64")
+    num_correct_chunks = helper.create_variable_for_type_inference(
+        dtype="int64")
 
     helper.append_op(
         type="chunk_eval",
@@ -1326,7 +1333,7 @@ def sequence_conv(input,
     filter_shape = [filter_size * input.shape[1], num_filters]
     filter_param = helper.create_parameter(
         attr=helper.param_attr, shape=filter_shape, dtype=dtype)
-    pre_bias = helper.create_tmp_variable(dtype)
+    pre_bias = helper.create_variable_for_type_inference(dtype)
 
     helper.append_op(
         type='sequence_conv',
@@ -1382,7 +1389,7 @@ def sequence_softmax(input, use_cudnn=False, name=None):
     """
     helper = LayerHelper('sequence_softmax', **locals())
     dtype = helper.input_dtype()
-    softmax_out = helper.create_tmp_variable(dtype)
+    softmax_out = helper.create_variable_for_type_inference(dtype)
     helper.append_op(
         type="sequence_softmax",
         inputs={"X": input},
@@ -1436,7 +1443,7 @@ def softmax(input, use_cudnn=True, name=None):
     """
     helper = LayerHelper('softmax', **locals())
     dtype = helper.input_dtype()
-    softmax_out = helper.create_tmp_variable(dtype)
+    softmax_out = helper.create_variable_for_type_inference(dtype)
     helper.append_op(
         type="softmax",
         inputs={"X": input},
@@ -1599,7 +1606,7 @@ def conv2d(input,
         dtype=dtype,
         default_initializer=_get_default_param_initializer())
 
-    pre_bias = helper.create_tmp_variable(dtype)
+    pre_bias = helper.create_variable_for_type_inference(dtype)
 
     helper.append_op(
         type=l_type,
@@ -1770,7 +1777,7 @@ def conv3d(input,
         dtype=dtype,
         default_initializer=_get_default_param_initializer())
 
-    pre_bias = helper.create_tmp_variable(dtype)
+    pre_bias = helper.create_variable_for_type_inference(dtype)
 
     helper.append_op(
         type=l_type,
@@ -1849,8 +1856,8 @@ def sequence_pool(input, pool_type):
     """
     helper = LayerHelper('sequence_pool', **locals())
     dtype = helper.input_dtype()
-    pool_out = helper.create_tmp_variable(dtype)
-    max_index = helper.create_tmp_variable(dtype)
+    pool_out = helper.create_variable_for_type_inference(dtype)
+    max_index = helper.create_variable_for_type_inference(dtype)
 
     helper.append_op(
         type="sequence_pool",
@@ -1886,7 +1893,7 @@ def sequence_concat(input, name=None):
            out = fluid.layers.sequence_concat(input=[seq1, seq2, seq3])
     """
     helper = LayerHelper('sequence_concat', **locals())
-    out = helper.create_tmp_variable(dtype=helper.input_dtype())
+    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
     helper.append_op(
         type='sequence_concat', inputs={'X': input}, outputs={'Out': [out]})
     return out
@@ -2013,7 +2020,7 @@ def sequence_slice(input, offset, length, name=None):
     """
     helper = LayerHelper("sequence_slice", **locals())
     dtype = helper.input_dtype()
-    out = helper.create_tmp_variable(dtype)
+    out = helper.create_variable_for_type_inference(dtype)
 
     offset.stop_gradient = True
     length.stop_gradient = True
@@ -2099,7 +2106,7 @@ def pool2d(input,
 
     helper = LayerHelper(l_type, **locals())
     dtype = helper.input_dtype()
-    pool_out = helper.create_tmp_variable(dtype)
+    pool_out = helper.create_variable_for_type_inference(dtype)
 
     helper.append_op(
         type=l_type,
@@ -2167,7 +2174,7 @@ def pool3d(input,
     l_type = "pool3d"
     helper = LayerHelper(l_type, **locals())
     dtype = helper.input_dtype()
-    pool_out = helper.create_tmp_variable(dtype)
+    pool_out = helper.create_variable_for_type_inference(dtype)
 
     helper.append_op(
         type=l_type,
@@ -2310,10 +2317,13 @@ def batch_norm(input,
     mean_out = mean
     # variance and variance out share the same memory
     variance_out = variance
-    saved_mean = helper.create_tmp_variable(dtype=dtype, stop_gradient=True)
-    saved_variance = helper.create_tmp_variable(dtype=dtype, stop_gradient=True)
+    saved_mean = helper.create_variable_for_type_inference(
+        dtype=dtype, stop_gradient=True)
+    saved_variance = helper.create_variable_for_type_inference(
+        dtype=dtype, stop_gradient=True)
 
-    batch_norm_out = input if in_place else helper.create_tmp_variable(dtype)
+    batch_norm_out = input if in_place else helper.create_variable_for_type_inference(
+        dtype)
 
     helper.append_op(
         type="batch_norm",
@@ -2430,9 +2440,11 @@ def layer_norm(input,
         inputs['Bias'] = bias
 
     # create output
-    mean_out = helper.create_tmp_variable(dtype=dtype, stop_gradient=True)
-    variance_out = helper.create_tmp_variable(dtype=dtype, stop_gradient=True)
-    layer_norm_out = helper.create_tmp_variable(dtype)
+    mean_out = helper.create_variable_for_type_inference(
+        dtype=dtype, stop_gradient=True)
+    variance_out = helper.create_variable_for_type_inference(
+        dtype=dtype, stop_gradient=True)
+    layer_norm_out = helper.create_variable_for_type_inference(dtype)
 
     helper.append_op(
         type="layer_norm",
@@ -2619,7 +2631,7 @@ def conv2d_transpose(input,
     img_filter = helper.create_parameter(
         dtype=input.dtype, shape=filter_shape, attr=helper.param_attr)
 
-    pre_bias = helper.create_tmp_variable(dtype=input.dtype)
+    pre_bias = helper.create_variable_for_type_inference(dtype=input.dtype)
     helper.append_op(
         type=op_type,
         inputs={'Input': [input],
@@ -2797,7 +2809,7 @@ def conv3d_transpose(input,
     img_filter = helper.create_parameter(
         dtype=input.dtype, shape=filter_shape, attr=helper.param_attr)
 
-    pre_bias = helper.create_tmp_variable(dtype=input.dtype)
+    pre_bias = helper.create_variable_for_type_inference(dtype=input.dtype)
     helper.append_op(
         type=l_type,
         inputs={'Input': [input],
@@ -2876,7 +2888,7 @@ def sequence_expand(x, y, ref_level=-1, name=None):
     """
     helper = LayerHelper('sequence_expand', input=x, **locals())
     dtype = helper.input_dtype()
-    tmp = helper.create_tmp_variable(dtype)
+    tmp = helper.create_variable_for_type_inference(dtype)
     helper.append_op(
         type='sequence_expand',
         inputs={'X': x,
@@ -2942,7 +2954,7 @@ def sequence_expand_as(x, y, name=None):
     """
     helper = LayerHelper('sequence_expand_as', input=x, **locals())
     dtype = helper.input_dtype()
-    tmp = helper.create_tmp_variable(dtype)
+    tmp = helper.create_variable_for_type_inference(dtype)
     helper.append_op(
         type='sequence_expand_as',
         inputs={'X': x,
@@ -2987,8 +2999,8 @@ def sequence_pad(x, pad_value, maxlen=None, name=None):
 
     helper = LayerHelper('sequence_pad', input=x, **locals())
     dtype = helper.input_dtype()
-    out = helper.create_tmp_variable(dtype)
-    length = helper.create_tmp_variable(dtype)
+    out = helper.create_variable_for_type_inference(dtype)
+    length = helper.create_variable_for_type_inference(dtype)
 
     pad_value.stop_gradient = True
     length.stop_gradient = True
@@ -3053,7 +3065,7 @@ def sequence_unpad(x, length, name=None):
 
     helper = LayerHelper('sequence_unpad', input=x, **locals())
     dtype = helper.input_dtype()
-    out = helper.create_tmp_variable(dtype)
+    out = helper.create_variable_for_type_inference(dtype)
 
     length.stop_gradient = True
 
@@ -3152,8 +3164,9 @@ def beam_search(pre_ids,
     score_type = scores.dtype
     id_type = ids.dtype
 
-    selected_scores = helper.create_tmp_variable(dtype=score_type)
-    selected_ids = helper.create_tmp_variable(dtype=id_type)
+    selected_scores = helper.create_variable_for_type_inference(
+        dtype=score_type)
+    selected_ids = helper.create_variable_for_type_inference(dtype=id_type)
 
     helper.append_op(
         type='beam_search',
@@ -3210,8 +3223,8 @@ def beam_search_decode(ids, scores, beam_size, end_id, name=None):
                 ids, scores, beam_size=5, end_id=0)
     """
     helper = LayerHelper('beam_search_decode', **locals())
-    sentence_ids = helper.create_tmp_variable(dtype=ids.dtype)
-    sentence_scores = helper.create_tmp_variable(dtype=ids.dtype)
+    sentence_ids = helper.create_variable_for_type_inference(dtype=ids.dtype)
+    sentence_scores = helper.create_variable_for_type_inference(dtype=ids.dtype)
 
     helper.append_op(
         type="beam_search_decode",
@@ -3341,8 +3354,8 @@ def lstm_unit(x_t,
                 param_attr=param_attr,
                 bias_attr=bias_attr)
     dtype = x_t.dtype
-    c = helper.create_tmp_variable(dtype)
-    h = helper.create_tmp_variable(dtype)
+    c = helper.create_variable_for_type_inference(dtype)
+    h = helper.create_variable_for_type_inference(dtype)
 
     helper.append_op(
         type='lstm_unit',
@@ -3396,7 +3409,7 @@ def reduce_sum(input, dim=None, keep_dim=False, name=None):
 
     """
     helper = LayerHelper('reduce_sum', **locals())
-    out = helper.create_tmp_variable(dtype=helper.input_dtype())
+    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
     if dim is not None and not isinstance(dim, list):
         dim = [dim]
     helper.append_op(
@@ -3453,7 +3466,7 @@ def reduce_mean(input, dim=None, keep_dim=False, name=None):
             fluid.layers.reduce_mean(x, dim=[0, 1]) # [4.0, 5.0]
     """
     helper = LayerHelper('reduce_mean', **locals())
-    out = helper.create_tmp_variable(dtype=helper.input_dtype())
+    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
     if dim is not None and not isinstance(dim, list):
         dim = [dim]
     helper.append_op(
@@ -3508,7 +3521,7 @@ def reduce_max(input, dim=None, keep_dim=False, name=None):
             fluid.layers.reduce_max(x, dim=[0, 1]) # [7.0, 8.0]
     """
     helper = LayerHelper('reduce_max', **locals())
-    out = helper.create_tmp_variable(dtype=helper.input_dtype())
+    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
     if dim is not None and not isinstance(dim, list):
         dim = [dim]
     helper.append_op(
@@ -3563,7 +3576,7 @@ def reduce_min(input, dim=None, keep_dim=False, name=None):
             fluid.layers.reduce_min(x, dim=[0, 1]) # [1.0, 2.0]
     """
     helper = LayerHelper('reduce_min', **locals())
-    out = helper.create_tmp_variable(dtype=helper.input_dtype())
+    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
     if dim is not None and not isinstance(dim, list):
         dim = [dim]
     helper.append_op(
@@ -3619,7 +3632,7 @@ def reduce_prod(input, dim=None, keep_dim=False, name=None):
             fluid.layers.reduce_prod(x, dim=[0, 1]) # [105.0, 384.0]
     """
     helper = LayerHelper('reduce_prod', **locals())
-    out = helper.create_tmp_variable(dtype=helper.input_dtype())
+    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
     if dim is not None and not isinstance(dim, list):
         dim = [dim]
     helper.append_op(
@@ -3679,7 +3692,7 @@ def split(input, num_or_sections, dim=-1, name=None):
             dim], 'len(num_or_sections) must not be more than input.shape[dim].'
         num = len(num_or_sections)
     outs = [
-        helper.create_tmp_variable(dtype=helper.input_dtype())
+        helper.create_variable_for_type_inference(dtype=helper.input_dtype())
         for i in range(num)
     ]
     helper.append_op(
@@ -3736,8 +3749,8 @@ def l2_normalize(x, axis, epsilon=1e-12, name=None):
         axis = 0
     helper = LayerHelper("l2_normalize", **locals())
 
-    out = helper.create_tmp_variable(dtype=x.dtype)
-    norm = helper.create_tmp_variable(dtype=x.dtype)
+    out = helper.create_variable_for_type_inference(dtype=x.dtype)
+    norm = helper.create_variable_for_type_inference(dtype=x.dtype)
     helper.append_op(
         type="norm",
         inputs={"X": x},
@@ -3846,7 +3859,7 @@ def matmul(x, y, transpose_x=False, transpose_y=False, alpha=1.0, name=None):
     __check_input(x, y)
 
     helper = LayerHelper('matmul', **locals())
-    out = helper.create_tmp_variable(dtype=x.dtype)
+    out = helper.create_variable_for_type_inference(dtype=x.dtype)
     helper.append_op(
         type='matmul',
         inputs={'X': x,
@@ -3917,8 +3930,8 @@ def topk(input, k, name=None):
             top5_values, top5_indices = layers.topk(input, k=5)
     """
     helper = LayerHelper("top_k", **locals())
-    values = helper.create_tmp_variable(dtype=input.dtype)
-    indices = helper.create_tmp_variable(dtype="int64")
+    values = helper.create_variable_for_type_inference(dtype=input.dtype)
+    indices = helper.create_variable_for_type_inference(dtype="int64")
     helper.append_op(
         type="top_k",
         inputs={"X": [input]},
@@ -3976,8 +3989,8 @@ def edit_distance(input, label, normalized=True, ignored_tokens=None):
 
     # remove some tokens from input and labels
     if ignored_tokens is not None and len(ignored_tokens) > 0:
-        erased_input = helper.create_tmp_variable(dtype="int64")
-        erased_label = helper.create_tmp_variable(dtype="int64")
+        erased_input = helper.create_variable_for_type_inference(dtype="int64")
+        erased_label = helper.create_variable_for_type_inference(dtype="int64")
 
         helper.append_op(
             type="sequence_erase",
@@ -3994,8 +4007,8 @@ def edit_distance(input, label, normalized=True, ignored_tokens=None):
         label = erased_label
 
     # edit distance op
-    edit_distance_out = helper.create_tmp_variable(dtype="int64")
-    sequence_num = helper.create_tmp_variable(dtype="int64")
+    edit_distance_out = helper.create_variable_for_type_inference(dtype="int64")
+    sequence_num = helper.create_variable_for_type_inference(dtype="int64")
     helper.append_op(
         type="edit_distance",
         inputs={"Hyps": [input],
@@ -4070,7 +4083,7 @@ def ctc_greedy_decoder(input, blank, name=None):
     _, topk_indices = topk(input, k=1)
 
     # ctc align op
-    ctc_out = helper.create_tmp_variable(dtype="int64")
+    ctc_out = helper.create_variable_for_type_inference(dtype="int64")
     helper.append_op(
         type="ctc_align",
         inputs={"Input": [topk_indices]},
@@ -4120,8 +4133,8 @@ def warpctc(input, label, blank=0, norm_by_times=False):
 
     """
     helper = LayerHelper('warpctc', **locals())
-    loss_out = helper.create_tmp_variable(dtype=input.dtype)
-    grad_out = helper.create_tmp_variable(dtype=input.dtype)
+    loss_out = helper.create_variable_for_type_inference(dtype=input.dtype)
+    grad_out = helper.create_variable_for_type_inference(dtype=input.dtype)
     helper.append_op(
         type='warpctc',
         inputs={'Logits': [input],
@@ -4182,7 +4195,7 @@ def sequence_reshape(input, new_dim):
             x_reshaped = fluid.layers.sequence_reshape(input=x, new_dim=10)
     """
     helper = LayerHelper('sequence_reshape', **locals())
-    out = helper.create_tmp_variable(helper.input_dtype())
+    out = helper.create_variable_for_type_inference(helper.input_dtype())
     helper.append_op(
         type='sequence_reshape',
         inputs={'X': [input]},
@@ -4279,9 +4292,9 @@ def nce(input,
             is_bias=True,
             dtype=input.dtype)
         inputs['Bias'] = b
-    cost = helper.create_tmp_variable(dtype=input.dtype)
-    sample_logits = helper.create_tmp_variable(dtype=input.dtype)
-    sample_labels = helper.create_tmp_variable(dtype=label.dtype)
+    cost = helper.create_variable_for_type_inference(dtype=input.dtype)
+    sample_logits = helper.create_variable_for_type_inference(dtype=input.dtype)
+    sample_labels = helper.create_variable_for_type_inference(dtype=label.dtype)
 
     if num_neg_samples is None:
         num_neg_samples = 10
@@ -4357,8 +4370,8 @@ def hsigmoid(input,
 
     helper = LayerHelper('hierarchical_sigmoid', **locals())
     dtype = helper.input_dtype()
-    out = helper.create_tmp_variable(dtype)
-    pre_out = helper.create_tmp_variable(dtype)
+    out = helper.create_variable_for_type_inference(dtype)
+    pre_out = helper.create_variable_for_type_inference(dtype)
     dim = input.shape[1]
     if num_classes < 2:
         raise ValueError("num_classes must not be less than 2.")
@@ -4418,8 +4431,8 @@ def transpose(x, perm, name=None):
                 (idx, perm[idx], len(x.shape)))
 
     helper = LayerHelper('transpose', **locals())
-    out = helper.create_tmp_variable(x.dtype)
-    x_shape = helper.create_tmp_variable(x.dtype)
+    out = helper.create_variable_for_type_inference(x.dtype)
+    x_shape = helper.create_variable_for_type_inference(x.dtype)
     helper.append_op(
         type='transpose2',
         inputs={'X': [x]},
@@ -4561,7 +4574,7 @@ def im2sequence(input,
         inputs["Y"] = input_image_size
         attrs["out_stride"] = out_stride
     helper = LayerHelper('im2sequence', **locals())
-    out = helper.create_tmp_variable(dtype=helper.input_dtype())
+    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
     helper.append_op(
         type='im2sequence', inputs=inputs, outputs={'Out': out}, attrs=attrs)
     return out
@@ -4594,7 +4607,7 @@ def row_conv(input, future_context_size, param_attr=None, act=None):
     filter_shape = [future_context_size + 1, input.shape[1]]
     filter_param = helper.create_parameter(
         attr=helper.param_attr, shape=filter_shape, dtype=dtype)
-    out = helper.create_tmp_variable(dtype)
+    out = helper.create_variable_for_type_inference(dtype)
     helper.append_op(
         type='row_conv',
         inputs={'X': [input],
@@ -4627,7 +4640,7 @@ def multiplex(inputs, index):
         raise ValueError("inputs should be a list object and contains at least "
                          "2 elements.")
 
-    out = helper.create_tmp_variable(inputs[0].dtype)
+    out = helper.create_variable_for_type_inference(inputs[0].dtype)
     helper.append_op(
         type='multiplex',
         inputs={'X': inputs,
@@ -4698,8 +4711,8 @@ def softmax_with_cross_entropy(logits,
                 logits=fc, label=label)
     """
     helper = LayerHelper('softmax_with_cross_entropy', **locals())
-    softmax = helper.create_tmp_variable(dtype=logits.dtype)
-    loss = helper.create_tmp_variable(dtype=logits.dtype)
+    softmax = helper.create_variable_for_type_inference(dtype=logits.dtype)
+    loss = helper.create_variable_for_type_inference(dtype=logits.dtype)
     helper.append_op(
         type='softmax_with_cross_entropy',
         inputs={'Logits': logits,
@@ -4749,8 +4762,8 @@ def smooth_l1(x, y, inside_weight=None, outside_weight=None, sigma=None):
     """
 
     helper = LayerHelper('smooth_l1_loss', **locals())
-    diff = helper.create_tmp_variable(dtype=x.dtype)
-    loss = helper.create_tmp_variable(dtype=x.dtype)
+    diff = helper.create_variable_for_type_inference(dtype=x.dtype)
+    loss = helper.create_variable_for_type_inference(dtype=x.dtype)
     helper.append_op(
         type='smooth_l1_loss',
         inputs={
@@ -4783,7 +4796,7 @@ def one_hot(input, depth):
             one_hot_label = layers.one_hot(input=label, depth=10)
     """
     helper = LayerHelper("one_hot", **locals())
-    one_hot_out = helper.create_tmp_variable(dtype='float32')
+    one_hot_out = helper.create_variable_for_type_inference(dtype='float32')
     helper.append_op(
         type="one_hot",
         inputs={'X': input},
@@ -4930,8 +4943,9 @@ def reshape(x, shape, actual_shape=None, act=None, inplace=False, name=None):
                 "except one unknown dimension.")
 
     helper = LayerHelper("reshape2", **locals())
-    x_shape = helper.create_tmp_variable(dtype=x.dtype)
-    out = x if inplace else helper.create_tmp_variable(dtype=x.dtype)
+    out = x if inplace else helper.create_variable_for_type_inference(
+        dtype=x.dtype)
+    x_shape = helper.create_variable_for_type_inference(dtype=x.dtype)
     helper.append_op(
         type="reshape2",
         inputs=inputs,
@@ -4980,8 +4994,8 @@ def squeeze(input, axes, name=None):
             y = layers.sequeeze(input=x, axes=[1])
     """
     helper = LayerHelper("squeeze", **locals())
-    out = helper.create_tmp_variable(dtype=input.dtype)
-    x_shape = helper.create_tmp_variable(dtype=input.dtype)
+    out = helper.create_variable_for_type_inference(dtype=input.dtype)
+    x_shape = helper.create_variable_for_type_inference(dtype=input.dtype)
     helper.append_op(
         type="squeeze2",
         inputs={"X": input},
@@ -5017,8 +5031,8 @@ def unsqueeze(input, axes, name=None):
             y = layers.unsequeeze(input=x, axes=[1])
     """
     helper = LayerHelper("unsqueeze", **locals())
-    out = helper.create_tmp_variable(dtype=input.dtype)
-    x_shape = helper.create_tmp_variable(dtype=input.dtype)
+    out = helper.create_variable_for_type_inference(dtype=input.dtype)
+    x_shape = helper.create_variable_for_type_inference(dtype=input.dtype)
     helper.append_op(
         type="unsqueeze2",
         inputs={"X": input},
@@ -5108,7 +5122,7 @@ def lod_reset(x, y=None, target_lod=None):
             out = layers.lod_reset(x=x, y=y)
     """
     helper = LayerHelper("lod_reset", **locals())
-    out = helper.create_tmp_variable(dtype=x.dtype)
+    out = helper.create_variable_for_type_inference(dtype=x.dtype)
     if y is not None:
         helper.append_op(
             type="lod_reset", inputs={'X': x,
@@ -5177,8 +5191,9 @@ def lrn(input, n=5, k=1.0, alpha=1e-4, beta=0.75, name=None):
             "dims of input must be 4(not %d), and it's order must be NCHW" %
             (dims))
 
-    mid_out = helper.create_tmp_variable(dtype=dtype, stop_gradient=True)
-    lrn_out = helper.create_tmp_variable(dtype)
+    mid_out = helper.create_variable_for_type_inference(
+        dtype=dtype, stop_gradient=True)
+    lrn_out = helper.create_variable_for_type_inference(dtype)
     helper.append_op(
         type="lrn",
         inputs={"X": input},
@@ -5243,7 +5258,7 @@ def pad(x, paddings, pad_value=0., name=None):
     """
     helper = LayerHelper('pad', input=x, **locals())
     dtype = helper.input_dtype()
-    out = helper.create_tmp_variable(dtype)
+    out = helper.create_variable_for_type_inference(dtype)
     helper.append_op(
         type='pad',
         inputs={'X': x},
@@ -5323,7 +5338,7 @@ def pad_constant_like(x, y, pad_value=0., name=None):
     """
     helper = LayerHelper('pad_constant_like', input=x, **locals())
     dtype = helper.input_dtype()
-    out = helper.create_tmp_variable(dtype)
+    out = helper.create_variable_for_type_inference(dtype)
     helper.append_op(
         type='pad_constant_like',
         inputs={'X': x,
@@ -5388,7 +5403,7 @@ def label_smooth(label,
         raise ValueError("The value of epsilon must be between 0 and 1.")
     helper = LayerHelper("label_smooth", **locals())
     label.stop_gradient = True
-    smooth_label = helper.create_tmp_variable(dtype)
+    smooth_label = helper.create_variable_for_type_inference(dtype)
     helper.append_op(
         type="label_smooth",
         inputs={"X": label,
@@ -5420,8 +5435,8 @@ def roi_pool(input, rois, pooled_height=1, pooled_width=1, spatial_scale=1.0):
     """
     helper = LayerHelper('roi_pool', **locals())
     dtype = helper.input_dtype()
-    pool_out = helper.create_tmp_variable(dtype)
-    argmaxes = helper.create_tmp_variable(dtype='int32')
+    pool_out = helper.create_variable_for_type_inference(dtype)
+    argmaxes = helper.create_variable_for_type_inference(dtype='int32')
     helper.append_op(
         type="roi_pool",
         inputs={"X": input,
@@ -5469,7 +5484,7 @@ def roi_align(input,
     """
     helper = LayerHelper('roi_align', **locals())
     dtype = helper.input_dtype()
-    align_out = helper.create_tmp_variable(dtype)
+    align_out = helper.create_variable_for_type_inference(dtype)
     helper.append_op(
         type="roi_align",
         inputs={"X": input,
@@ -5594,7 +5609,7 @@ def image_resize(input,
         out_h = int(input.shape[2] * scale)
         out_w = int(input.shape[3] * scale)
 
-    out = helper.create_tmp_variable(dtype)
+    out = helper.create_variable_for_type_inference(dtype)
     helper.append_op(
         type=resample_methods[resample],
         inputs=inputs,
@@ -5703,7 +5718,7 @@ def gather(input, index):
     """
     helper = LayerHelper('gather', **locals())
     dtype = helper.input_dtype()
-    out = helper.create_tmp_variable(dtype)
+    out = helper.create_variable_for_type_inference(dtype)
     helper.append_op(
         type="gather",
         inputs={"X": input,
@@ -5743,7 +5758,7 @@ def scatter(input, index, updates, name=None):
     """
     helper = LayerHelper('scatter', **locals())
     dtype = helper.input_dtype()
-    out = helper.create_tmp_variable(dtype)
+    out = helper.create_variable_for_type_inference(dtype)
     helper.append_op(
         type="scatter",
         inputs={"X": input,
@@ -5803,7 +5818,7 @@ def sequence_scatter(input, index, updates, name=None):
     """
     helper = LayerHelper('sequence_scatter', **locals())
     dtype = helper.input_dtype()
-    out = helper.create_tmp_variable(dtype)
+    out = helper.create_variable_for_type_inference(dtype)
     helper.append_op(
         type="sequence_scatter",
         inputs={"X": input,
@@ -5833,7 +5848,7 @@ def random_crop(x, shape, seed=None):
     """
     helper = LayerHelper("random_crop", **locals())
     dtype = x.dtype
-    out = helper.create_tmp_variable(dtype)
+    out = helper.create_variable_for_type_inference(dtype)
     if seed is None:
         seed = np.random.randint(-65536, 65536)
     op_attrs = {"shape": shape}
@@ -5879,7 +5894,7 @@ def log(x, name=None):
     """
     helper = LayerHelper('log', **locals())
     dtype = helper.input_dtype(input_param_name='x')
-    out = helper.create_tmp_variable(dtype)
+    out = helper.create_variable_for_type_inference(dtype)
     helper.append_op(type="log", inputs={"X": x}, outputs={"Out": out})
     return out
 
@@ -5910,7 +5925,7 @@ def relu(x, name=None):
     """
     helper = LayerHelper('relu', **locals())
     dtype = helper.input_dtype(input_param_name='x')
-    out = helper.create_tmp_variable(dtype)
+    out = helper.create_variable_for_type_inference(dtype)
     helper.append_op(type="relu", inputs={"X": x}, outputs={"Out": out})
     return out
 
@@ -5949,9 +5964,9 @@ def mean_iou(input, label, num_classes):
     """
     helper = LayerHelper('mean_iou', **locals())
     dtype = helper.input_dtype()
-    out_mean_iou = helper.create_tmp_variable(dtype='float32')
-    out_wrong = helper.create_tmp_variable(dtype='int32')
-    out_correct = helper.create_tmp_variable(dtype='int32')
+    out_mean_iou = helper.create_variable_for_type_inference(dtype='float32')
+    out_wrong = helper.create_variable_for_type_inference(dtype='int32')
+    out_correct = helper.create_variable_for_type_inference(dtype='int32')
     helper.append_op(
         type="mean_iou",
         inputs={"Predictions": input,
@@ -6043,7 +6058,7 @@ def crop(x, shape=None, offsets=None, name=None):
     if offsets is None:
         offsets = [0] * len(x.shape)
 
-    out = helper.create_tmp_variable(x.dtype)
+    out = helper.create_variable_for_type_inference(x.dtype)
     ipts = {'X': x}
     attrs = {}
     if isinstance(shape, Variable):
@@ -6123,7 +6138,7 @@ def rank_loss(label, left, right, name=None):
     if not (isinstance(right, Variable)):
         raise ValueError("The right should be a Variable")
 
-    out = helper.create_tmp_variable("float32")
+    out = helper.create_variable_for_type_inference("float32")
 
     helper.append_op(
         type='rank_loss',
@@ -6169,8 +6184,8 @@ def margin_rank_loss(label, left, right, margin=0.1, name=None):
         raise ValueError("The left should be a Variable.")
     if not isinstance(right, Variable):
         raise ValueError("The right should be a Variable.")
-    out = helper.create_tmp_variable(left.dtype)
-    act = helper.create_tmp_variable(left.dtype)
+    out = helper.create_variable_for_type_inference(left.dtype)
+    act = helper.create_variable_for_type_inference(left.dtype)
     helper.append_op(
         type='margin_rank_loss',
         inputs={"Label": label,
@@ -6255,7 +6270,7 @@ def pad2d(input,
 
     helper = LayerHelper('pad2d', **locals())
     dtype = helper.input_dtype(input_param_name='input')
-    out = helper.create_tmp_variable(dtype)
+    out = helper.create_variable_for_type_inference(dtype)
     helper.append_op(
         type='pad2d',
         inputs={'X': input},
@@ -6284,7 +6299,7 @@ def elu(x, alpha=1.0, name=None):
         output(${out_type}): ${out_comment}
     """
     helper = LayerHelper('elu', **locals())
-    out = helper.create_tmp_variable(dtype=x.dtype)
+    out = helper.create_variable_for_type_inference(dtype=x.dtype)
     helper.append_op(
         type='elu',
         inputs={'X': x},
@@ -6307,7 +6322,7 @@ def relu6(x, threshold=6.0, name=None):
         output(${out_type}): ${out_comment}
     """
     helper = LayerHelper('relu6', **locals())
-    out = helper.create_tmp_variable(dtype=x.dtype)
+    out = helper.create_variable_for_type_inference(dtype=x.dtype)
     helper.append_op(
         type='relu6',
         inputs={'X': x},
@@ -6330,7 +6345,7 @@ def pow(x, factor=1.0, name=None):
         output(${out_type}): ${out_comment}
     """
     helper = LayerHelper('pow', **locals())
-    out = helper.create_tmp_variable(dtype=x.dtype)
+    out = helper.create_variable_for_type_inference(dtype=x.dtype)
     helper.append_op(
         type='pow',
         inputs={'X': x},
@@ -6354,7 +6369,7 @@ def stanh(x, scale_a=2.0 / 3.0, scale_b=1.7159, name=None):
         output(${out_type}): ${out_comment}
     """
     helper = LayerHelper('stanh', **locals())
-    out = helper.create_tmp_variable(dtype=x.dtype)
+    out = helper.create_variable_for_type_inference(dtype=x.dtype)
     helper.append_op(
         type='stanh',
         inputs={'X': x},
@@ -6379,7 +6394,7 @@ def hard_sigmoid(x, slope=0.2, offset=0.5, name=None):
         output(${out_type}): ${out_comment}
     """
     helper = LayerHelper('hard_sigmoid', **locals())
-    out = helper.create_tmp_variable(dtype=x.dtype)
+    out = helper.create_variable_for_type_inference(dtype=x.dtype)
     helper.append_op(
         type='hard_sigmoid',
         inputs={'X': x},
@@ -6403,7 +6418,7 @@ def swish(x, beta=1.0, name=None):
         output(${out_type}): ${out_comment}
     """
     helper = LayerHelper('swish', **locals())
-    out = helper.create_tmp_variable(dtype=x.dtype)
+    out = helper.create_variable_for_type_inference(dtype=x.dtype)
     helper.append_op(
         type='swish',
         inputs={'X': x},
@@ -6455,7 +6470,7 @@ def prelu(x, mode, param_attr=None, name=None):
         dtype='float32',
         is_bias=False,
         default_initializer=Constant(1.0))
-    out = helper.create_tmp_variable(dtype)
+    out = helper.create_variable_for_type_inference(dtype)
     helper.append_op(
         type="prelu",
         inputs={"X": x,
@@ -6479,7 +6494,7 @@ def brelu(x, t_min=0.0, t_max=24.0, name=None):
         output(${out_type}): ${out_comment}
     """
     helper = LayerHelper('brelu', **locals())
-    out = helper.create_tmp_variable(dtype=x.dtype)
+    out = helper.create_variable_for_type_inference(dtype=x.dtype)
     helper.append_op(
         type='brelu',
         inputs={'X': x},
@@ -6502,7 +6517,7 @@ def leaky_relu(x, alpha=0.02, name=None):
         output(${out_type}): ${out_comment}
     """
     helper = LayerHelper('leaky_relu', **locals())
-    out = helper.create_tmp_variable(dtype=x.dtype)
+    out = helper.create_variable_for_type_inference(dtype=x.dtype)
     helper.append_op(
         type='leaky_relu',
         inputs={'X': x},
@@ -6524,7 +6539,7 @@ def soft_relu(x, threshold=40.0, name=None):
         output(${out_type}): ${out_comment}
     """
     helper = LayerHelper('soft_relu', **locals())
-    out = helper.create_tmp_variable(dtype=x.dtype)
+    out = helper.create_variable_for_type_inference(dtype=x.dtype)
     helper.append_op(
         type='soft_relu',
         inputs={'X': x},
@@ -6591,8 +6606,8 @@ def flatten(x, axis=1, name=None):
     if not (isinstance(axis, int)) or axis > len(x.shape) or axis < 0:
         raise ValueError("The axis should be a int, and in range [0, rank(x)]")
 
-    out = helper.create_tmp_variable(x.dtype)
-    x_shape = helper.create_tmp_variable(x.dtype)
+    out = helper.create_variable_for_type_inference(x.dtype)
+    x_shape = helper.create_variable_for_type_inference(x.dtype)
     helper.append_op(
         type='flatten2',
         inputs={"X": x},
@@ -6638,7 +6653,8 @@ def sequence_enumerate(input, win_size, pad_value=0, name=None):
             out = fluid.layers.sequence_enumerate(input=x, win_size=3, pad_value=0)
     """
     helper = LayerHelper('sequence_enumerate', **locals())
-    out = helper.create_tmp_variable(helper.input_dtype(), stop_gradient=True)
+    out = helper.create_variable_for_type_inference(
+        helper.input_dtype(), stop_gradient=True)
     helper.append_op(
         type='sequence_enumerate',
         inputs={'X': input},
@@ -6678,9 +6694,9 @@ def sequence_mask(x, maxlen=None, dtype='int64', name=None):
 
     helper = LayerHelper('sequence_mask', **locals())
     if name is None:
-        out = helper.create_tmp_variable(dtype=dtype)
+        out = helper.create_variable_for_type_inference(dtype=dtype)
     else:
-        out = helper.create_tmp_variable(dtype=dtype, name=name)
+        out = helper.create_variable_for_type_inference(dtype=dtype, name=name)
 
     helper.append_op(
         type='sequence_mask',
@@ -6723,7 +6739,7 @@ def stack(x, axis=0):
     if not isinstance(x, list) and not isinstance(x, tuple):
         x = [x]
 
-    out = helper.create_tmp_variable(x[0].dtype)
+    out = helper.create_variable_for_type_inference(x[0].dtype)
     helper.append_op(
         type='stack', inputs={'X': x}, outputs={'Y': out},
         attrs={'axis': axis})
@@ -6761,7 +6777,7 @@ def unstack(x, axis=0, num=None):
 
     outs = []
     for _ in num:
-        outs.append(helper.create_tmp_variable(x.dtype))
+        outs.append(helper.create_variable_for_type_inference(x.dtype))
 
     helper.append_op(
         type='unstack',
@@ -6813,7 +6829,7 @@ def expand(x, expand_times, name=None):
     """
     helper = LayerHelper('expand', input=x, **locals())
     dtype = helper.input_dtype(input_param_name='x')
-    out = helper.create_tmp_variable(dtype)
+    out = helper.create_variable_for_type_inference(dtype)
     helper.append_op(
         type='expand',
         inputs={'X': x},
@@ -6852,7 +6868,7 @@ def uniform_random_batch_size_like(input,
     """
 
     helper = LayerHelper('uniform_random_batch_size_like', **locals())
-    out = helper.create_tmp_variable(dtype)
+    out = helper.create_variable_for_type_inference(dtype)
     c_dtype = convert_np_dtype_to_dtype_(dtype)
     helper.append_op(
         type='uniform_random_batch_size_like',
@@ -6889,7 +6905,7 @@ def gaussian_random(shape, mean=0.0, std=1.0, seed=0, dtype='float32'):
     """
 
     helper = LayerHelper('gaussian_random', **locals())
-    out = helper.create_tmp_variable(dtype)
+    out = helper.create_variable_for_type_inference(dtype)
     c_dtype = convert_np_dtype_to_dtype_(dtype)
     helper.append_op(
         type='gaussian_random',
@@ -6924,7 +6940,7 @@ def sampling_id(x, min=0.0, max=1.0, seed=0, dtype='float32'):
     """
 
     helper = LayerHelper('sampling_id', **locals())
-    out = helper.create_tmp_variable(dtype)
+    out = helper.create_variable_for_type_inference(dtype)
     helper.append_op(
         type='sampling_id',
         inputs={'X': x},
@@ -6963,7 +6979,7 @@ def gaussian_random_batch_size_like(input,
     """
 
     helper = LayerHelper('gaussian_random_batch_size_like', **locals())
-    out = helper.create_tmp_variable(dtype)
+    out = helper.create_variable_for_type_inference(dtype)
     c_dtype = convert_np_dtype_to_dtype_(dtype)
     helper.append_op(
         type='gaussian_random_batch_size_like',
@@ -6995,7 +7011,8 @@ def sum(x):
     """
 
     helper = LayerHelper('sum', **locals())
-    out = helper.create_tmp_variable(dtype=helper.input_dtype('x'))
+    out = helper.create_variable_for_type_inference(
+        dtype=helper.input_dtype('x'))
     helper.append_op(
         type='sum',
         inputs={'X': x},
@@ -7022,7 +7039,8 @@ def slice(input, axes, starts, ends):
     """
 
     helper = LayerHelper('slice', **locals())
-    out = helper.create_tmp_variable(dtype=helper.input_dtype('input'))
+    out = helper.create_variable_for_type_inference(
+        dtype=helper.input_dtype('input'))
     helper.append_op(
         type='slice',
         inputs={'Input': input},
@@ -7048,7 +7066,8 @@ def shape(input):
     """
 
     helper = LayerHelper('shape', **locals())
-    out = helper.create_tmp_variable(dtype=helper.input_dtype('input'))
+    out = helper.create_variable_for_type_inference(
+        dtype=helper.input_dtype('input'))
     helper.append_op(
         type='shape', inputs={'Input': input}, outputs={'Out': out})
 
@@ -7065,7 +7084,7 @@ def _elementwise_op(helper):
     use_mkldnn = helper.kwargs.get('use_mkldnn', False)
     name = helper.kwargs.get('name', None)
     if name is None:
-        out = helper.create_tmp_variable(dtype=x.dtype)
+        out = helper.create_variable_for_type_inference(dtype=x.dtype)
     else:
         out = helper.create_variable(
             name=name, dtype=x.dtype, persistable=False)
@@ -7099,7 +7118,7 @@ def scale(x, scale=1.0, bias=0.0, bias_after_scale=True, act=None, name=None):
 
     helper = LayerHelper('scale', **locals())
     if name is None:
-        out = helper.create_tmp_variable(dtype=x.dtype)
+        out = helper.create_variable_for_type_inference(dtype=x.dtype)
     else:
         out = helper.create_variable(
             name=name, dtype=x.dtype, persistable=False)
@@ -7165,7 +7184,7 @@ def _logical_op(op_name, x, y, out=None, name=None, binary_op=True):
 
     if out is None:
         if name is None:
-            out = helper.create_tmp_variable(dtype=x.dtype)
+            out = helper.create_variable_for_type_inference(dtype=x.dtype)
         else:
             out = helper.create_variable(
                 name=name, dtype=x.dtype, persistable=False)
@@ -7273,7 +7292,7 @@ def clip(x, min, max, name=None):
     helper = LayerHelper("clip", **locals())
 
     if name is None:
-        out = helper.create_tmp_variable(dtype=x.dtype)
+        out = helper.create_variable_for_type_inference(dtype=x.dtype)
     else:
         out = helper.create_variable(
             name=name, dtype=x.dtype, persistable=False)
@@ -7305,7 +7324,7 @@ def clip_by_norm(x, max_norm, name=None):
     helper = LayerHelper("clip_by_norm", **locals())
 
     if name is None:
-        out = helper.create_tmp_variable(dtype=x.dtype)
+        out = helper.create_variable_for_type_inference(dtype=x.dtype)
     else:
         out = helper.create_variable(
             name=name, dtype=x.dtype, persistable=False)
@@ -7335,7 +7354,7 @@ def mean(x, name=None):
     helper = LayerHelper("mean", **locals())
 
     if name is None:
-        out = helper.create_tmp_variable(dtype=x.dtype)
+        out = helper.create_variable_for_type_inference(dtype=x.dtype)
     else:
         out = helper.create_variable(
             name=name, dtype=x.dtype, persistable=False)
@@ -7365,7 +7384,7 @@ def mul(x, y, x_num_col_dims=1, y_num_col_dims=1, name=None):
     helper = LayerHelper("mul", **locals())
 
     if name is None:
-        out = helper.create_tmp_variable(dtype=x.dtype)
+        out = helper.create_variable_for_type_inference(dtype=x.dtype)
     else:
         out = helper.create_variable(
             name=name, dtype=x.dtype, persistable=False)
@@ -7399,7 +7418,7 @@ def sigmoid_cross_entropy_with_logits(x, label, name=None):
     helper = LayerHelper("sigmoid_cross_entropy_with_logits", **locals())
 
     if name is None:
-        out = helper.create_tmp_variable(dtype=x.dtype)
+        out = helper.create_variable_for_type_inference(dtype=x.dtype)
     else:
         out = helper.create_variable(
             name=name, dtype=x.dtype, persistable=False)
@@ -7429,7 +7448,7 @@ def maxout(x, groups, name=None):
     helper = LayerHelper("maxout", **locals())
 
     if name is None:
-        out = helper.create_tmp_variable(dtype=x.dtype)
+        out = helper.create_variable_for_type_inference(dtype=x.dtype)
     else:
         out = helper.create_variable(
             name=name, dtype=x.dtype, persistable=False)
@@ -7468,7 +7487,7 @@ def affine_channel(x, scale=None, bias=None, data_layout='NCHW', name=None):
     helper = LayerHelper("affine_channel", **locals())
 
     if name is None:
-        out = helper.create_tmp_variable(dtype=x.dtype)
+        out = helper.create_variable_for_type_inference(dtype=x.dtype)
     else:
         out = helper.create_variable(
             name=name, dtype=x.dtype, persistable=False)
diff --git a/python/paddle/fluid/layers/tensor.py b/python/paddle/fluid/layers/tensor.py
index 9c6a2112a69637aad4baff9a0296330620e3c52f..09a7cb8dc9339afa666f8cf09e92a27ffba8a9b3 100644
--- a/python/paddle/fluid/layers/tensor.py
+++ b/python/paddle/fluid/layers/tensor.py
@@ -152,7 +152,7 @@ def cast(x, dtype):
             result = fluid.layers.cast(x=data, dtype='float64')
     """
     helper = LayerHelper('cast', **locals())
-    out = helper.create_tmp_variable(dtype=dtype)
+    out = helper.create_variable_for_type_inference(dtype=dtype)
     helper.append_op(
         type='cast',
         inputs={'X': [x]},
@@ -184,7 +184,7 @@ def concat(input, axis=0, name=None):
            out = fluid.layers.concat(input=[Efirst, Esecond, Ethird, Efourth])
     """
     helper = LayerHelper('concat', **locals())
-    out = helper.create_tmp_variable(dtype=helper.input_dtype())
+    out = helper.create_variable_for_type_inference(dtype=helper.input_dtype())
     helper.append_op(
         type='concat',
         inputs={'X': input},
@@ -221,7 +221,8 @@ def sums(input, out=None):
     """
     helper = LayerHelper('sum', **locals())
     if out is None:
-        out = helper.create_tmp_variable(dtype=helper.input_dtype())
+        out = helper.create_variable_for_type_inference(
+            dtype=helper.input_dtype())
     helper.append_op(
         type='sum',
         inputs={'X': input},
@@ -252,7 +253,7 @@ def assign(input, output=None):
     """
     helper = LayerHelper('assign', **locals())
     if output is None:
-        output = helper.create_tmp_variable(dtype=input.dtype)
+        output = helper.create_variable_for_type_inference(dtype=input.dtype)
     if isinstance(input, Variable):
         helper.append_op(
             type='assign', inputs={'X': [input]}, outputs={'Out': [output]})
@@ -311,7 +312,7 @@ def fill_constant(shape, dtype, value, force_cpu=False, out=None):
 
     helper = LayerHelper("fill_constant", **locals())
     if out is None:
-        out = helper.create_tmp_variable(dtype=dtype)
+        out = helper.create_variable_for_type_inference(dtype=dtype)
     helper.append_op(
         type='fill_constant',
         inputs={},
@@ -358,7 +359,7 @@ def fill_constant_batch_size_like(input,
         ${out_comment}.
     """
     helper = LayerHelper("fill_constant_batch_size_like", **locals())
-    out = helper.create_tmp_variable(dtype=dtype)
+    out = helper.create_variable_for_type_inference(dtype=dtype)
     helper.append_op(
         type='fill_constant_batch_size_like',
         inputs={'Input': input},
@@ -396,7 +397,7 @@ def argmin(x, axis=0):
           out = fluid.layers.argmin(x=in, axis=-1)
     """
     helper = LayerHelper("arg_min", **locals())
-    out = helper.create_tmp_variable(VarDesc.VarType.INT64)
+    out = helper.create_variable_for_type_inference(VarDesc.VarType.INT64)
     helper.append_op(
         type='arg_min',
         inputs={'X': x},
@@ -427,7 +428,7 @@ def argmax(x, axis=0):
           out = fluid.layers.argmax(x=in, axis=-1)
     """
     helper = LayerHelper("arg_max", **locals())
-    out = helper.create_tmp_variable(VarDesc.VarType.INT64)
+    out = helper.create_variable_for_type_inference(VarDesc.VarType.INT64)
     helper.append_op(
         type='arg_max',
         inputs={'X': x},
@@ -477,8 +478,10 @@ def argsort(input, axis=-1, name=None):
             out, indices = fluid.layers.argsort(input, axis=0)
     """
     helper = LayerHelper("argsort", **locals())
-    out = helper.create_tmp_variable(dtype=input.dtype, stop_gradient=True)
-    ids = helper.create_tmp_variable(VarDesc.VarType.INT64, stop_gradient=True)
+    out = helper.create_variable_for_type_inference(
+        dtype=input.dtype, stop_gradient=True)
+    ids = helper.create_variable_for_type_inference(
+        VarDesc.VarType.INT64, stop_gradient=True)
     helper.append_op(
         type='argsort',
         inputs={'X': input},
@@ -562,7 +565,7 @@ def reverse(x, axis):
     if isinstance(axis, int):
         axis = [axis]
     helper = LayerHelper("reverse", **locals())
-    out = helper.create_tmp_variable(dtype=x.dtype)
+    out = helper.create_variable_for_type_inference(dtype=x.dtype)
     helper.append_op(
         type='reverse',
         inputs={'Input': x},
@@ -654,7 +657,7 @@ def has_inf(x):
         Variable: The tensor variable storing the output, only a bool value.
     """
     helper = LayerHelper("isinf", **locals())
-    out = helper.create_tmp_variable(dtype=x.dtype)
+    out = helper.create_variable_for_type_inference(dtype=x.dtype)
     helper.append_op(type="isinf", inputs={"X": x}, outputs={"Out": out})
     return out
 
@@ -670,7 +673,7 @@ def has_nan(x):
         Variable: The tensor variable storing the output, only a bool value.
     """
     helper = LayerHelper("isnan", **locals())
-    out = helper.create_tmp_variable(dtype=x.dtype)
+    out = helper.create_variable_for_type_inference(dtype=x.dtype)
     helper.append_op(type="isnan", inputs={"X": x}, outputs={"Out": out})
     return out
 
@@ -687,6 +690,6 @@ def isfinite(x):
         Variable: The tensor variable storing the output, contains a bool value.
     """
     helper = LayerHelper("isfinite", **locals())
-    out = helper.create_tmp_variable(dtype=x.dtype)
+    out = helper.create_variable_for_type_inference(dtype=x.dtype)
     helper.append_op(type="isfinite", inputs={"X": x}, outputs={"Out": out})
     return out
diff --git a/python/paddle/fluid/regularizer.py b/python/paddle/fluid/regularizer.py
index 97644df007117be35ccd4d66495d8ca0ec7b0e88..c151fbd17208bb6e3104e8d0f6590392c6095987 100644
--- a/python/paddle/fluid/regularizer.py
+++ b/python/paddle/fluid/regularizer.py
@@ -151,7 +151,7 @@ class L2DecayRegularizer(WeightDecayRegularizer):
             decay = block.create_var(
                 dtype="float32",
                 shape=param.shape,
-                type=core.VarDesc.VarType.SELECTED_ROWS)
+                type=core.VarDesc.VarType.LOD_TENSOR)
             block.append_op(
                 type='extract_rows', inputs={'X': grad}, outputs={'Out': idx})
             block.append_op(
@@ -228,7 +228,7 @@ class L1DecayRegularizer(WeightDecayRegularizer):
             decay = block.create_var(
                 dtype="float32",
                 shape=param.shape,
-                type=core.VarDesc.VarType.SELECTED_ROWS)
+                type=core.VarDesc.VarType.LOD_TENSOR)
             block.append_op(
                 type='extract_rows', inputs={'X': grad}, outputs={'Out': idx})
             block.append_op(
diff --git a/python/paddle/fluid/tests/unittests/test_slice_var.py b/python/paddle/fluid/tests/unittests/test_slice_var.py
index fab63b7d5631829feffd26fc1dce2bd338d2036b..b16c744603534ad07f9318a5a195f1d7550b1f10 100644
--- a/python/paddle/fluid/tests/unittests/test_slice_var.py
+++ b/python/paddle/fluid/tests/unittests/test_slice_var.py
@@ -30,7 +30,6 @@ class TestSliceVar(unittest.TestCase):
             var = program.global_block().create_var(
                 name=str(random.randint(10000, 99999)),
                 persistable=True,
-                # dtype=core.VarDesc.VarType.LOD_TENSOR,
                 shape=shape)
             var_list.append(var)
         blocks = slice_variable(var_list, 10, min_size)