diff --git a/paddle/framework/operator.h b/paddle/framework/operator.h
index b7c9c39402d57daf0aec97d98535ac8a8d9c0150..28a253ec0b0ef1d8e0ee4068692479469c687242 100644
--- a/paddle/framework/operator.h
+++ b/paddle/framework/operator.h
@@ -336,6 +336,14 @@ class InferShapeContext {
     return &var->Get<Tensor>();
   }
 
+  void ShareLoD(const std::string& in, const std::string& out) const {
+    PADDLE_ENFORCE(InputVar(in)->IsType<LoDTensor>(),
+                   "The Input(%s) must be LoDTensor.", in);
+    PADDLE_ENFORCE(OutputVar(out)->IsType<LoDTensor>(),
+                   "The Output(%s) must be LoDTensor.", out);
+    Output<LoDTensor>(out)->set_lod(Input<LoDTensor>(in)->lod());
+  }
+
  private:
   const OperatorBase& op_;
   const Scope& scope_;
diff --git a/paddle/operators/accuracy_op.cc b/paddle/operators/accuracy_op.cc
index 0c813748b2989a8f0c00a359345747242dd21dd8..32479ae5a35194e942fb4012c268c534200e1db6 100644
--- a/paddle/operators/accuracy_op.cc
+++ b/paddle/operators/accuracy_op.cc
@@ -40,6 +40,7 @@ class AccuracyOp : public framework::OperatorWithKernel {
                       "inference size must be the same as label size");
 
     ctx.Output<framework::LoDTensor>("Accuracy")->Resize({1});
+    ctx.ShareLoD("Inference", "Accuracy");
   }
 };
 
@@ -58,7 +59,11 @@ class AccuracyOpMaker : public framework::OpProtoAndCheckerMaker {
         R"DOC(Accuracy. It will print accuracy rate for classification.
 The accuracy is:
 ..  math::
-accuracy = \\frac{NumOfCorrectPredicts}{NumOfAllSamples})DOC");
+accuracy = \\frac{NumOfCorrectPredicts}{NumOfAllSamples})
+
+Both the input `Inference` and `Label` can carry the LoD (Level of Details)
+information, or not. But the output only shares the LoD with input `Inference`.
+DOC");
   }
 };
 
diff --git a/paddle/operators/cos_sim_op.cc b/paddle/operators/cos_sim_op.cc
index 72c446493684246959656dc048e7f0e761665423..840848fa087e991a7ba7fb9c11643a0d770343ac 100644
--- a/paddle/operators/cos_sim_op.cc
+++ b/paddle/operators/cos_sim_op.cc
@@ -57,6 +57,7 @@ class CosSimOp : public framework::OperatorWithKernel {
     ctx.Output<framework::LoDTensor>("Out")->Resize({x_dims[0], 1});
     ctx.Output<framework::LoDTensor>("XNorm")->Resize({x_dims[0], 1});
     ctx.Output<framework::LoDTensor>("YNorm")->Resize({y_dims[0], 1});
+    ctx.ShareLoD("X", "Out");
   }
 };
 
@@ -81,10 +82,13 @@ Cosine Similarity Operator.
 
 The equation is: Out = X^T * Y / (sqrt(X^T * X) * sqrt(Y^T * Y)).
 
-Input(X) and Input(Y) must have the same shape, except that the 1st dimension
-of Input(Y) could be just 1 (different from Input(X)), which will be
-broadcasted to match the shape of Input(X) before computing their cosine
+The input `X` and `Y` must have the same shape, except that the 1st dimension
+of input `Y` could be just 1 (different from input `X`), which will be
+broadcasted to match the shape of input `X` before computing their cosine
 similarity.
+
+Both the input `X` and `Y` can carry the LoD (Level of Details) information,
+or not. But the output only shares the LoD with input `X`.
 )DOC");
   }
 };
diff --git a/paddle/operators/elementwise_mul_op.cc b/paddle/operators/elementwise_mul_op.cc
index ee6e975b443691bf71cec904565ced20406f3fba..304e45fa5ba6e507aa76aa0aaaf33d4ef9469725 100644
--- a/paddle/operators/elementwise_mul_op.cc
+++ b/paddle/operators/elementwise_mul_op.cc
@@ -38,6 +38,7 @@ class ElementWiseMulOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_GE(x_dim.size(), y_dim.size(),
                       "Rank of first input must >= rank of second input.")
     ctx.Output<framework::LoDTensor>("Out")->Resize(x_dim);
+    ctx.ShareLoD("X", "Out");
   }
 };
 
@@ -63,11 +64,15 @@ Limited elementwise multiple operator.The equation is: Out = X ⊙ Y.
 2. Y's shape is a subset of X. 
    Y will be broadcasted to match the shape of X and axis should be dimension index Y in X.
    example:
+
       shape(X) = (2, 3, 4, 5), shape(Y) = (,)
       shape(X) = (2, 3, 4, 5), shape(Y) = (5,)
       shape(X) = (2, 3, 4, 5), shape(Y) = (4, 5)
       shape(X) = (2, 3, 4, 5), shape(Y) = (3, 4), with axis=1
       shape(X) = (2, 3, 4, 5), shape(Y) = (2), with axis=0
+
+Both the input X and Y can carry the LoD (Level of Details) information,
+or not. But the output only shares the LoD with input X.
 )DOC");
   }
 };
diff --git a/paddle/operators/fc_op.cc b/paddle/operators/fc_op.cc
index e5d0f3c3724262a60a463ef3beadd9906d3ebaf6..56fe654d1e8df9ce2978dde4eb653bcfe38ebd02 100644
--- a/paddle/operators/fc_op.cc
+++ b/paddle/operators/fc_op.cc
@@ -186,6 +186,10 @@ W_i is a 2-D matrix of size (K x N), where N means the number of neurons
 in the fully connected layer. B is a 1-D vector of size N.
 Thus, the output Out is a 2-D matrix of size (M x N).
 Activation type can be set to `identity` (default), `sigmoid` or `softmax`.
+
+All the inputs can carry the LoD (Level of Details) information,
+or not. But the output only shares the LoD with first input (`X[0]`).
+)DOC");
 )DOC");
   }
 };
diff --git a/paddle/operators/fill_zeros_like_op.cc b/paddle/operators/fill_zeros_like_op.cc
index ba7857cc65f6860a6156674c6addc2bfdce21a99..a238b59b7877b2b13d439d130170931372c46c61 100644
--- a/paddle/operators/fill_zeros_like_op.cc
+++ b/paddle/operators/fill_zeros_like_op.cc
@@ -23,15 +23,14 @@ class FillZerosLikeOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE_NOT_NULL(
-        ctx.InputVar("Src"),
-        "Input(Src) of FillZerosLikeOp should not be null.");
-    PADDLE_ENFORCE_NOT_NULL(
-        ctx.OutputVar("Dst"),
-        "Output(Dst) of FillZerosLikeOp should not be null.");
-
-    ctx.Output<framework::LoDTensor>("Dst")->Resize(
-        ctx.Input<framework::Tensor>("Src")->dims());
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input(X) of FillZerosLikeOp should not be null.");
+    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Y"),
+                            "Output(Y) of FillZerosLikeOp should not be null.");
+
+    ctx.Output<framework::LoDTensor>("Y")->Resize(
+        ctx.Input<framework::Tensor>("X")->dims());
+    ctx.ShareLoD("X", "Y");
   }
 };
 
@@ -40,8 +39,8 @@ class FillZerosLikeOpMaker : public framework::OpProtoAndCheckerMaker {
   FillZerosLikeOpMaker(framework::OpProto *proto,
                        framework::OpAttrChecker *op_checker)
       : framework::OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("Src", "The input of fill-zeros-like op.");
-    AddOutput("Dst", "The varibale will be filled up with zeros.");
+    AddInput("X", "The input of fill-zeros-like op.");
+    AddOutput("Y", "The varibale will be filled up with zeros.");
     AddComment(R"DOC(
 Fill up a vriable with zeros.
 
diff --git a/paddle/operators/fill_zeros_like_op.h b/paddle/operators/fill_zeros_like_op.h
index 969998ce2eae02b8ad057c6259703e51559bf98a..4474581784531faee1741f0b143743e31cc3788f 100644
--- a/paddle/operators/fill_zeros_like_op.h
+++ b/paddle/operators/fill_zeros_like_op.h
@@ -23,7 +23,7 @@ template <typename Place, typename T>
 class FillZerosLikeKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
-    auto* output = context.Output<framework::Tensor>("Dst");
+    auto* output = context.Output<framework::Tensor>("Y");
     output->mutable_data<T>(context.GetPlace());
     auto t = framework::EigenVector<T>::Flatten(*output);
     t.device(context.GetEigenDevice<Place>()) = t.constant(static_cast<T>(0));
diff --git a/paddle/operators/lookup_table_op.cc b/paddle/operators/lookup_table_op.cc
index 07f6dfabca5879e3de6004e59d2e87f7fa68d66c..8f533f1cc3c887c4101428545a9196a054f3da1b 100644
--- a/paddle/operators/lookup_table_op.cc
+++ b/paddle/operators/lookup_table_op.cc
@@ -35,6 +35,7 @@ class LookupTableOp : public framework::OperatorWithKernel {
     auto output_t = ctx.Output<framework::LoDTensor>("Out");
 
     output_t->Resize({ids_t->dims()[0], table_t->dims()[1]});
+    ctx.ShareLoD("Ids", "Out");
   }
 };
 
@@ -50,9 +51,13 @@ class LookupTableOpMaker : public framework::OpProtoAndCheckerMaker {
              "An input with type int32 or int64"
              "contains the ids to be looked up in W.");
     AddOutput("Out", "The lookup results, which have the same type with W.");
-    AddComment(
-        "This operator is used to perform lookups on the parameter W,"
-        "then concatenated into a dense tensor.");
+    AddComment(R"DOC(
+This operator is used to perform lookups on the parameter W,
+then concatenated into a dense tensor.
+
+The input `Ids` can carry the LoD (Level of Details) information,
+or not. And the output only shares the LoD with input `Ids`.
+)DOC");
   }
 };
 
diff --git a/paddle/operators/mean_op.cc b/paddle/operators/mean_op.cc
index 7d7eeb59a23435036dc33c1e4fe6dd1c4a1a2f62..96540ff45495103557ef8e95eee5dad5d91643da 100644
--- a/paddle/operators/mean_op.cc
+++ b/paddle/operators/mean_op.cc
@@ -37,7 +37,8 @@ class MeanOpMaker : public framework::OpProtoAndCheckerMaker {
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "The input of mean op");
     AddOutput("Out", "The output of mean op").NotInGradient();
-    AddComment("Mean Operator");
+    AddComment(R"DOC( Mean Operator
+)DOC");
   }
 };
 
diff --git a/paddle/operators/minus_op.cc b/paddle/operators/minus_op.cc
index a97bbecdca1779df330d1053cf359bb658aa75c2..5036f9f98ad36327aa1b03cbc8cbabdc4fdb7bb7 100644
--- a/paddle/operators/minus_op.cc
+++ b/paddle/operators/minus_op.cc
@@ -41,6 +41,7 @@ class MinusOp : public framework::OperatorWithKernel {
         left_tensor->numel(), right_tensor->numel(),
         "Minus operator must take two tensor with same num of elements");
     ctx.Output<framework::LoDTensor>("Out")->Resize(left_tensor->dims());
+    ctx.ShareLoD("X", "Out");
   }
 };
 
@@ -54,7 +55,12 @@ class MinusOpMaker : public framework::OpProtoAndCheckerMaker {
 
     AddComment(R"DOC(Minus Operator
 
-Equation: Out = X - Y
+Equation:
+
+    Out = X - Y
+
+Both the input `X` and `Y` can carry the LoD (Level of Details) information,
+or not. But the output only shares the LoD with input `X`.
 )DOC");
   }
 };
diff --git a/paddle/operators/mul_op.cc b/paddle/operators/mul_op.cc
index b6d320b415e02549e85cb36ab517b0b5433887d5..b2409a18707860f656cffa2fc743f3fa80e1ec66 100644
--- a/paddle/operators/mul_op.cc
+++ b/paddle/operators/mul_op.cc
@@ -55,6 +55,7 @@ class MulOp : public framework::OperatorWithKernel {
         "First matrix's width must be equal with second matrix's height.");
     ctx.Output<framework::LoDTensor>("Out")->Resize(
         {x_mat_dims[0], y_mat_dims[1]});
+    ctx.ShareLoD("X", "Out");
   }
 };
 
@@ -83,9 +84,14 @@ class MulOpMaker : public framework::OpProtoAndCheckerMaker {
         .SetDefault(1)
         .EqualGreaterThan(1);
     AddComment(R"DOC(
-Two Element Mul Operator.
+Mul operator is used to perform matrix multiplication for input X and Y.
 
-The equation is: Out = X * Y
+The equation is:
+
+    Out = X * Y
+
+Both the input `X` and `Y` can carry the LoD (Level of Details) information,
+or not. But the output only shares the LoD with input `X`.
 )DOC");
   }
 };
diff --git a/paddle/operators/onehot_cross_entropy_op.cc b/paddle/operators/onehot_cross_entropy_op.cc
index f38be3549f3c5d2443f61739fc32cdca74197649..1d87032d275dbef8c9cf6569c897dbbfce46bd16 100644
--- a/paddle/operators/onehot_cross_entropy_op.cc
+++ b/paddle/operators/onehot_cross_entropy_op.cc
@@ -40,6 +40,7 @@ class OnehotCrossEntropyOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_EQ(label->dims().size(), 1, "label's dimension must be 1.");
     PADDLE_ENFORCE_EQ(X->dims()[0], label->dims()[0]);
     ctx.Output<framework::LoDTensor>("Y")->Resize({X->dims()[0], 1});
+    ctx.ShareLoD("X", "Y");
   }
 };
 
@@ -69,6 +70,8 @@ OnehotCrossEntropy Operator.
 
                 Y[i] = -log(X[i][j])
 
+Both the input `X` and `Label` can carry the LoD (Level of Details) information,
+or not. But the output only shares the LoD with input `X`.
 )DOC");
   }
 };
diff --git a/paddle/operators/prelu_op.cc b/paddle/operators/prelu_op.cc
index 7ae80b296850f2f433c89d904ebf32355b2a29c7..2b7b82a3e1bf8c6c544bd3226770d8d98a5c7f7a 100644
--- a/paddle/operators/prelu_op.cc
+++ b/paddle/operators/prelu_op.cc
@@ -38,6 +38,7 @@ class PReluOp : public framework::OperatorWithKernel {
                             "Output(Out) should not be null");
     auto *out = ctx.Output<framework::LoDTensor>("Out");
     out->Resize(in->dims());
+    ctx.ShareLoD("X", "Out");
   }
 };
 
@@ -55,6 +56,8 @@ The equation is:
   f(x) = alpha * x , for x < 0
   f(x) = x         , for x >= 0
 
+The input `X` can carry the LoD (Level of Details) information,
+or not. And the output shares the LoD with input `X`.
 )DOC");
   }
 };
diff --git a/paddle/operators/rowwise_add_op.cc b/paddle/operators/rowwise_add_op.cc
index 2a3fd3be941d91aaa6b014df91d3025f07767577..90cdb2558bf11401f023ebb7ca836f0d1a20b17c 100644
--- a/paddle/operators/rowwise_add_op.cc
+++ b/paddle/operators/rowwise_add_op.cc
@@ -45,6 +45,7 @@ class RowwiseAddOp : public framework::OperatorWithKernel {
         "The width of two operands must be same");
     PADDLE_ENFORCE_EQ(ctx.OutputSize("Out"), 1, "The output size must be 1");
     ctx.Output<framework::LoDTensor>("Out")->Resize(x_dims);
+    ctx.ShareLoD("X", "Out");
   }
 };
 
diff --git a/paddle/operators/scale_op.cc b/paddle/operators/scale_op.cc
index d1f42e8662537d35e17429f9d436fdc0e5a1dc11..ca1bc4ac80551037e251daa5f8f52ee11ba5b81e 100644
--- a/paddle/operators/scale_op.cc
+++ b/paddle/operators/scale_op.cc
@@ -35,6 +35,7 @@ class ScaleOp : public framework::OperatorWithKernel {
     auto *in = ctx.Input<framework::Tensor>("X");
     auto *out = ctx.Output<framework::LoDTensor>("Out");
     out->Resize(in->dims());
+    ctx.ShareLoD("X", "Out");
   }
 };
 
diff --git a/paddle/operators/sigmoid_op.cc b/paddle/operators/sigmoid_op.cc
index 992b19965e0ca9ce7dba1b8b3c5b7780af06eb45..42befa22d0cbf5158da368be76482dd3e66f9e44 100644
--- a/paddle/operators/sigmoid_op.cc
+++ b/paddle/operators/sigmoid_op.cc
@@ -30,6 +30,7 @@ class SigmoidOp : public framework::OperatorWithKernel {
 
     ctx.Output<framework::LoDTensor>("Y")->Resize(
         ctx.Input<Tensor>("X")->dims());
+    ctx.ShareLoD("X", "Y");
   }
 };
 
diff --git a/paddle/operators/squared_l2_distance_op.cc b/paddle/operators/squared_l2_distance_op.cc
index 39f4305877de20d451bc35fe698a0eabf9758d57..dfe8e6decd33a123711c350fadfc94cd1838b40b 100644
--- a/paddle/operators/squared_l2_distance_op.cc
+++ b/paddle/operators/squared_l2_distance_op.cc
@@ -57,6 +57,7 @@ class SquaredL2DistanceOp : public framework::OperatorWithKernel {
     ctx.Output<framework::LoDTensor>("sub_result")
         ->Resize({x_dims[0], x->numel() / x_dims[0]});
     ctx.Output<framework::LoDTensor>("Out")->Resize({x_dims[0], 1});
+    ctx.ShareLoD("X", "Out");
   }
 };
 
@@ -79,6 +80,9 @@ class SquaredL2DistanceOpMaker : public framework::OpProtoAndCheckerMaker {
     input or to 1. If the first dimension of target is 1, SquaredL2DistanceOp
     will broadcast target's first dimension to input's first dimension.
     You can decide whether calculate the gradient of input and target.
+
+    Both the input X and Y can carry the LoD (Level of Details) information,
+    or not. But the output only shares the LoD with input X.
     )DOC");
   }
 };
diff --git a/paddle/operators/sum_op.cc b/paddle/operators/sum_op.cc
index 41e05c27f9029b2664685d3979fadcfd2bf6dbce..ebc57d6b7bb7dcb28f4b387ec4f03841948a6458 100644
--- a/paddle/operators/sum_op.cc
+++ b/paddle/operators/sum_op.cc
@@ -39,6 +39,7 @@ class SumOp : public framework::OperatorWithKernel {
       PADDLE_ENFORCE(in_dim == dim, "Input tensors must have same shape");
     }
     out->Resize(in_dim);
+    ctx.ShareLoD(ctx.op().Inputs("X")[0], "Out");
   }
 };
 
@@ -49,8 +50,11 @@ class SumOpMaker : public framework::OpProtoAndCheckerMaker {
     AddInput("X", "the input tensors of sum operator.").AsDuplicable();
     AddOutput("Out", "the output tensor of sum operator.");
     AddComment(R"DOC(
-            Sum the input tensors.
-        )DOC");
+Sum the input tensors.
+
+All the inputs can carry the LoD (Level of Details) information,
+or not. But the output only shares the LoD with the first input.
+)DOC");
   }
 };
 
diff --git a/python/paddle/v2/framework/tests/test_fill_zeros_like_op.py b/python/paddle/v2/framework/tests/test_fill_zeros_like_op.py
index 2473daaba24438819f3f55ccc40fe1c64ee59960..eff8fa87d9c0dafc6935604101e94ee6c8b081ce 100644
--- a/python/paddle/v2/framework/tests/test_fill_zeros_like_op.py
+++ b/python/paddle/v2/framework/tests/test_fill_zeros_like_op.py
@@ -6,8 +6,8 @@ from op_test import OpTest
 class TestFillZerosLikeOp(OpTest):
     def setUp(self):
         self.op_type = "fill_zeros_like"
-        self.inputs = {'Src': np.random.random((219, 232)).astype("float32")}
-        self.outputs = {'Dst': np.zeros_like(self.inputs["Src"])}
+        self.inputs = {'X': np.random.random((219, 232)).astype("float32")}
+        self.outputs = {'Y': np.zeros_like(self.inputs["X"])}
 
     def test_check_output(self):
         self.check_output()