diff --git a/paddle/framework/operator_test.cc b/paddle/framework/operator_test.cc
index 8b4bb01a7bb80eaccee40f14fa82617505b1e2e5..7f0ec90adef7881a3a324b1ac652fa65f8a8b8d2 100644
--- a/paddle/framework/operator_test.cc
+++ b/paddle/framework/operator_test.cc
@@ -116,10 +116,13 @@ class OpWithKernelTest : public OperatorWithKernel {
 
  protected:
   void InferShape(framework::InferShapeContextBase* ctx) const override {}
+  DataType IndicateDataType(const ExecutionContext& ctx) const override {
+    return DataType::FP32;
+  }
 };
 
 template <typename T1, typename T2>
-class CPUKernelTest : public OpKernel {
+class CPUKernelTest : public OpKernel<float> {
  public:
   void Compute(const ExecutionContext& ctx) const {
     std::cout << "this is cpu kernel" << std::endl;
@@ -146,7 +149,7 @@ class OpKernelTestMultiInputsProtoAndCheckerMaker
   }
 };
 
-class CPUKernalMultiInputsTest : public OpKernel {
+class CPUKernalMultiInputsTest : public OpKernel<float> {
  public:
   void Compute(const ExecutionContext& ctx) const {
     auto xs = ctx.op().Inputs("xs");
diff --git a/paddle/operators/gather_op.cc b/paddle/operators/gather_op.cc
index 0e3cd174adee1e50d0a63861286a26d325484efb..da22bd0c52c27d7decd10e2e2b34fa38d0620da8 100644
--- a/paddle/operators/gather_op.cc
+++ b/paddle/operators/gather_op.cc
@@ -37,6 +37,11 @@ class GatherOp : public framework::OperatorWithKernel {
     output_dims[0] = batch_size;
     ctx->SetOutputDim("Out", output_dims);
   }
+
+  framework::DataType IndicateDataType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::ToDataType(ctx.Input<Tensor>("X")->type());
+  }
 };
 
 class GatherGradOp : public framework::OperatorWithKernel {
@@ -47,6 +52,11 @@ class GatherGradOp : public framework::OperatorWithKernel {
   void InferShape(framework::InferShapeContextBase* ctx) const override {
     ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X"));
   }
+
+  framework::DataType IndicateDataType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::ToDataType(ctx.Input<Tensor>("X")->type());
+  }
 };
 
 class GatherOpMaker : public framework::OpProtoAndCheckerMaker {
diff --git a/paddle/operators/gaussian_random_op.cc b/paddle/operators/gaussian_random_op.cc
index fc340c181cc6c54190dde19a67f020ac5ac13d07..5cd2c7d2c066cd31e2d38a3c0d682f02339b4d59 100644
--- a/paddle/operators/gaussian_random_op.cc
+++ b/paddle/operators/gaussian_random_op.cc
@@ -56,6 +56,11 @@ class GaussianRandomOp : public framework::OperatorWithKernel {
                    "dims can be one int or array. dims must be set.");
     ctx->SetOutputDim("Out", framework::make_ddim(temp));
   }
+
+  framework::DataType IndicateDataType(
+      const framework::ExecutionContext& ctx) const override {
+    return static_cast<framework::DataType>(Attr<int>("data_type"));
+  }
 };
 
 class GaussianRandomOpMaker : public framework::OpProtoAndCheckerMaker {
@@ -76,6 +81,8 @@ Use to initialize tensor with gaussian random generator.
                  "Random seed of generator."
                  "0 means use system wide seed")
         .SetDefault(0);
+    AddAttr<int>("data_type", "output data type")
+        .SetDefault(framework::DataType::FP32);
   }
 };
 
diff --git a/paddle/operators/lookup_table_op.cc b/paddle/operators/lookup_table_op.cc
index 9b1314bfbade8551d98b0fbabb7c2968d7600db5..929008fbcbe03bd6591b0a02252b343c46d00b8f 100644
--- a/paddle/operators/lookup_table_op.cc
+++ b/paddle/operators/lookup_table_op.cc
@@ -36,6 +36,11 @@ class LookupTableOp : public framework::OperatorWithKernel {
     ctx->SetOutputDim("Out", {ids_dims[0], table_dims[1]});
     ctx->ShareLoD("Ids", /*->*/ "Out");
   }
+
+  framework::DataType IndicateDataType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::ToDataType(ctx.Input<Tensor>("W")->type());
+  }
 };
 
 class LookupTableOpMaker : public framework::OpProtoAndCheckerMaker {
@@ -69,6 +74,11 @@ class LookupTableOpGrad : public framework::OperatorWithKernel {
     auto table_dims = ctx->GetInputDim("W");
     ctx->SetOutputDim(framework::GradVarName("W"), table_dims);
   }
+
+  framework::DataType IndicateDataType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::ToDataType(ctx.Input<Tensor>("W")->type());
+  }
 };
 
 }  // namespace operators
diff --git a/paddle/operators/multiplex_op.cc b/paddle/operators/multiplex_op.cc
index 9896d269ccc86d8fdc3bf6375e44ef5bf3e6b9c7..a069127a19a1d0ba4eaa2b3450a1c46262ace3ed 100644
--- a/paddle/operators/multiplex_op.cc
+++ b/paddle/operators/multiplex_op.cc
@@ -50,6 +50,11 @@ class MultiplexOp : public framework::OperatorWithKernel {
     }
     ctx->SetOutputDim("Out", in_dim);
   }
+
+  framework::DataType IndicateDataType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::ToDataType(ctx.MultiInput<Tensor>("X")[0]->type());
+  }
 };
 
 class MultiplexOpMaker : public framework::OpProtoAndCheckerMaker {
@@ -99,6 +104,11 @@ class MultiplexGradOp : public framework::OperatorWithKernel {
     }
     ctx->SetOutputsDim(framework::GradVarName("X"), d_ins);
   }
+
+  framework::DataType IndicateDataType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::ToDataType(ctx.MultiInput<Tensor>("X")[0]->type());
+  }
 };
 
 }  // namespace operators
diff --git a/paddle/operators/scatter_op.cc b/paddle/operators/scatter_op.cc
index 3fc4a39ebc5526bfed61ba667c3cdc214cdd056c..619acfc8b62ed8e50dd60d110280d981f5c1e3ab 100644
--- a/paddle/operators/scatter_op.cc
+++ b/paddle/operators/scatter_op.cc
@@ -48,6 +48,11 @@ class ScatterOp : public framework::OperatorWithKernel {
     }
     ctx->SetOutputDim("Out", ref_dims);
   }
+
+  framework::DataType IndicateDataType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::ToDataType(ctx.Input<Tensor>("X")->type());
+  }
 };
 
 class ScatterGradOp : public framework::OperatorWithKernel {
@@ -60,6 +65,11 @@ class ScatterGradOp : public framework::OperatorWithKernel {
                       ctx->GetInputDim("Updates"));
     ctx->SetOutputDim(framework::GradVarName("Ref"), ctx->GetInputDim("Ref"));
   }
+
+  framework::DataType IndicateDataType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::ToDataType(ctx.Input<Tensor>("X")->type());
+  }
 };
 
 class ScatterOpMaker : public framework::OpProtoAndCheckerMaker {
diff --git a/paddle/operators/softmax_with_cross_entropy_op.cc b/paddle/operators/softmax_with_cross_entropy_op.cc
index e2299b254458cdd42dee4683561d4d5c81653fb1..de7c532421c48927c77309db4fc713f9b5209967 100644
--- a/paddle/operators/softmax_with_cross_entropy_op.cc
+++ b/paddle/operators/softmax_with_cross_entropy_op.cc
@@ -13,6 +13,7 @@
    limitations under the License. */
 
 #include "paddle/operators/softmax_with_cross_entropy_op.h"
+#include <paddle/function/TensorType.h>
 
 namespace paddle {
 namespace operators {
@@ -115,6 +116,11 @@ class SoftmaxWithCrossEntropyOp : public framework::OperatorWithKernel {
     ctx->ShareLoD("Logits", /*->*/ "Softmax");
     ctx->ShareLoD("Logits", /*->*/ "Loss");
   }
+
+  framework::DataType IndicateDataType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::ToDataType(ctx.Input<Tensor>("Logits")->type());
+  }
 };
 
 class SoftmaxWithCrossEntropyOpGrad : public framework::OperatorWithKernel {
@@ -149,6 +155,11 @@ class SoftmaxWithCrossEntropyOpGrad : public framework::OperatorWithKernel {
     ctx->SetOutputDim(framework::GradVarName("Logits"),
                       ctx->GetInputDim("Softmax"));
   }
+
+  framework::DataType IndicateDataType(
+      const framework::ExecutionContext& ctx) const override {
+    return framework::ToDataType(ctx.Input<Tensor>("Logits")->type());
+  }
 };
 
 }  // namespace operators
diff --git a/paddle/operators/uniform_random_op.cc b/paddle/operators/uniform_random_op.cc
index 878d71802abb3c855b20293713282894e3b32416..97b1d0bed4595cb750e4d2122f294f10edfbe0ff 100644
--- a/paddle/operators/uniform_random_op.cc
+++ b/paddle/operators/uniform_random_op.cc
@@ -62,6 +62,11 @@ class UniformRandomOp : public framework::OperatorWithKernel {
     }
     ctx->SetOutputDim("Out", framework::make_ddim(temp));
   }
+
+  framework::DataType IndicateDataType(
+      const framework::ExecutionContext& ctx) const override {
+    return static_cast<framework::DataType>(Attr<int>("data_type"));
+  }
 };
 
 class UniformRandomOpMaker : public framework::OpProtoAndCheckerMaker {
@@ -80,6 +85,8 @@ Used to initialize tensor with uniform random generator.
                  "Random seed of uniform random. "
                  "0 means generate a seed by system")
         .SetDefault(0);
+    AddAttr<int>("data_type", "output tensor data type")
+        .SetDefault(framework::DataType::FP32);
   }
 };
 }  // namespace operators