diff --git a/paddle/fluid/framework/attribute.cc b/paddle/fluid/framework/attribute.cc
index 0dcecb62dba971b48c4f11c0ef47494be40eeea0..fabf2abfc803b8838edb48aa01ab8896799c97ac 100644
--- a/paddle/fluid/framework/attribute.cc
+++ b/paddle/fluid/framework/attribute.cc
@@ -64,6 +64,13 @@ Attribute GetAttrValue(const proto::OpDesc::Attr& attr_desc) {
     case proto::AttrType::LONG: {
       return attr_desc.l();
     }
+    case proto::AttrType::LONGS: {
+      std::vector<int64_t> val(attr_desc.longs_size());
+      for (int i = 0; i < attr_desc.longs_size(); ++i) {
+        val[i] = attr_desc.longs(i);
+      }
+      return val;
+    }
     default:
       PADDLE_THROW("Unsupport attr type %d", attr_desc.type());
   }
diff --git a/paddle/fluid/framework/attribute.h b/paddle/fluid/framework/attribute.h
index 14ca3e96209ed17f12e87fda8506806514698977..d9c76881b7e98d0b7cd29024b98c8f7720398c66 100644
--- a/paddle/fluid/framework/attribute.h
+++ b/paddle/fluid/framework/attribute.h
@@ -26,6 +26,113 @@ limitations under the License. */
 
 namespace paddle {
 namespace framework {
+
+template <typename T>
+struct ExtractAttribute {
+  explicit ExtractAttribute(const std::string& attr_name)
+      : attr_name_(attr_name) {}
+
+  T* operator()(Attribute& attr) const {
+    T* attr_value = nullptr;
+    try {
+      attr_value = &boost::get<T>(attr);
+    } catch (boost::bad_get& bad_get) {
+      PADDLE_THROW("Cannot get attribute %s by type %s, its type is %s",
+                   attr_name_, paddle::platform::demangle(typeid(T).name()),
+                   paddle::platform::demangle(attr.type().name()));
+    }
+    return attr_value;
+  }
+
+  const std::string& attr_name_;
+};
+
+// special handle bool
+// FIXME(yuyang18): Currently we cast bool into int in python binding. It is
+// hard to change the logic there. In another way, we should correct handle
+// if the user set `some_flag=1`.
+//
+// FIX ME anytime if there is a better solution.
+template <>
+struct ExtractAttribute<bool> {
+  explicit ExtractAttribute(const std::string& attr_name)
+      : attr_name_(attr_name) {}
+
+  bool* operator()(Attribute& attr) const {
+    if (attr.type() == typeid(int)) {  // NOLINT
+      int val = boost::get<int>(attr);
+      attr = static_cast<bool>(val);
+    } else if (attr.type() == typeid(float)) {  // NOLINT
+      float val = boost::get<float>(attr);
+      attr = static_cast<bool>(val);
+    }
+    bool* attr_value = nullptr;
+    try {
+      attr_value = &boost::get<bool>(attr);
+    } catch (boost::bad_get& bad_get) {
+      PADDLE_THROW("Cannot get attribute %s by type bool, its type is %s",
+                   attr_name_, paddle::platform::demangle(attr.type().name()));
+    }
+    return attr_value;
+  }
+
+  const std::string& attr_name_;
+};
+
+template <>
+struct ExtractAttribute<int64_t> {
+  explicit ExtractAttribute(const std::string& attr_name)
+      : attr_name_(attr_name) {}
+
+  int64_t* operator()(Attribute& attr) const {
+    if (attr.type() == typeid(int)) {  // NOLINT
+      int val = boost::get<int>(attr);
+      attr = static_cast<int64_t>(val);
+    } else if (attr.type() == typeid(float)) {  // NOLINT
+      int val = boost::get<float>(attr);
+      attr = static_cast<int64_t>(val);
+    }
+    int64_t* attr_value = nullptr;
+    try {
+      attr_value = &boost::get<int64_t>(attr);
+    } catch (boost::bad_get& bad_get) {
+      PADDLE_THROW("Cannot get attribute %s by type int64_t, its type is %s",
+                   attr_name_, paddle::platform::demangle(attr.type().name()));
+    }
+    return attr_value;
+  }
+
+  const std::string& attr_name_;
+};
+
+template <>
+struct ExtractAttribute<std::vector<int64_t>> {
+  explicit ExtractAttribute(const std::string& attr_name)
+      : attr_name_(attr_name) {}
+
+  std::vector<int64_t>* operator()(Attribute& attr) const {
+    if (attr.type() == typeid(std::vector<int>)) {  // NOLINT
+      std::vector<int> val = boost::get<std::vector<int>>(attr);
+      std::vector<int64_t> vec(val.begin(), val.end());
+      attr = vec;
+    } else if (attr.type() == typeid(std::vector<float>)) {  // NOLINT
+      std::vector<float> val = boost::get<std::vector<float>>(attr);
+      std::vector<int64_t> vec(val.begin(), val.end());
+      attr = vec;
+    }
+    std::vector<int64_t>* attr_value = nullptr;
+    try {
+      attr_value = &boost::get<std::vector<int64_t>>(attr);
+    } catch (boost::bad_get& bad_get) {
+      PADDLE_THROW("Cannot get attribute %s by type int64_t, its type is %s",
+                   attr_name_, paddle::platform::demangle(attr.type().name()));
+    }
+    return attr_value;
+  }
+
+  const std::string& attr_name_;
+};
+
 template <typename T>
 inline proto::AttrType AttrTypeID() {
   Attribute tmp = T();
@@ -42,7 +149,11 @@ class AttrReader {
   inline const T& Get(const std::string& name) const {
     PADDLE_ENFORCE(attrs_.count(name) != 0, "%s should be in AttributeMap",
                    name);
-    return boost::get<T>(attrs_.at(name));
+
+    Attribute& attr = const_cast<Attribute&>(attrs_.at(name));
+    ExtractAttribute<T> extract_attr(name);
+    T* attr_value = extract_attr(attr);
+    return *attr_value;
   }
 
  private:
@@ -82,7 +193,7 @@ class DefaultValueSetter {
  public:
   explicit DefaultValueSetter(T default_value)
       : default_value_(default_value) {}
-  void operator()(T& value) const { value = default_value_; }
+  void operator()(T& value) const { value = default_value_; }  // NOLINT
 
  private:
   T default_value_;
@@ -117,84 +228,6 @@ class EnumInContainer {
   std::unordered_set<T> container_;
 };
 
-template <typename T>
-struct ExtractAttribute {
-  explicit ExtractAttribute(const std::string& attr_name)
-      : attr_name_(attr_name) {}
-
-  T* operator()(Attribute& attr) const {
-    T* attr_value = nullptr;
-    try {
-      attr_value = &boost::get<T>(attr);
-    } catch (boost::bad_get& bad_get) {
-      PADDLE_THROW("Cannot get attribute %s by type %s, its type is %s",
-                   attr_name_, paddle::platform::demangle(typeid(T).name()),
-                   paddle::platform::demangle(attr.type().name()));
-    }
-    return attr_value;
-  }
-
-  const std::string& attr_name_;
-};
-
-// special handle bool
-// FIXME(yuyang18): Currently we cast bool into int in python binding. It is
-// hard to change the logic there. In another way, we should correct handle
-// if the user set `some_flag=1`.
-//
-// FIX ME anytime if there is a better solution.
-template <>
-struct ExtractAttribute<bool> {
-  explicit ExtractAttribute(const std::string& attr_name)
-      : attr_name_(attr_name) {}
-
-  bool* operator()(Attribute& attr) const {
-    if (attr.type() == typeid(int)) {  // NOLINT
-      int val = boost::get<int>(attr);
-      attr = static_cast<bool>(val);
-    } else if (attr.type() == typeid(float)) {  // NOLINT
-      float val = boost::get<float>(attr);
-      attr = static_cast<bool>(val);
-    }
-    bool* attr_value = nullptr;
-    try {
-      attr_value = &boost::get<bool>(attr);
-    } catch (boost::bad_get& bad_get) {
-      PADDLE_THROW("Cannot get attribute %s by type bool, its type is %s",
-                   attr_name_, paddle::platform::demangle(attr.type().name()));
-    }
-    return attr_value;
-  }
-
-  const std::string& attr_name_;
-};
-
-template <>
-struct ExtractAttribute<int64_t> {
-  explicit ExtractAttribute(const std::string& attr_name)
-      : attr_name_(attr_name) {}
-
-  int64_t* operator()(Attribute& attr) const {
-    if (attr.type() == typeid(int)) {  // NOLINT
-      int val = boost::get<int>(attr);
-      attr = static_cast<int64_t>(val);
-    } else if (attr.type() == typeid(float)) {  // NOLINT
-      int val = boost::get<float>(attr);
-      attr = static_cast<int64_t>(val);
-    }
-    int64_t* attr_value = nullptr;
-    try {
-      attr_value = &boost::get<int64_t>(attr);
-    } catch (boost::bad_get& bad_get) {
-      PADDLE_THROW("Cannot get attribute %s by type int64_t, its type is %s",
-                   attr_name_, paddle::platform::demangle(attr.type().name()));
-    }
-    return attr_value;
-  }
-
-  const std::string& attr_name_;
-};
-
 // check whether a certain attribute fit its limits
 // an attribute can have more than one limits
 template <typename T>
@@ -235,7 +268,7 @@ class TypedAttrChecker {
     return *this;
   }
 
-  void operator()(AttributeMap& attr_map) const {
+  void operator()(AttributeMap& attr_map) const {  // NOLINT
     if (!attr_map.count(attr_name_)) {
       // user do not set this attr
       PADDLE_ENFORCE(!default_value_setter_.empty(),
@@ -271,7 +304,7 @@ class OpAttrChecker {
     return *(checker.target<TypedAttrChecker<T>>());
   }
 
-  void Check(AttributeMap& attr_map) const {
+  void Check(AttributeMap& attr_map) const {  // NOLINT
     for (const auto& checker : attr_checkers_) {
       checker(attr_map);
     }
diff --git a/paddle/fluid/framework/details/multi_devices_graph_pass.cc b/paddle/fluid/framework/details/multi_devices_graph_pass.cc
index 4f3889a71e865ce6f9799d976c6e6ec5c365429b..ebd1d644bcce0554904ee0931827ef43a6d52b11 100644
--- a/paddle/fluid/framework/details/multi_devices_graph_pass.cc
+++ b/paddle/fluid/framework/details/multi_devices_graph_pass.cc
@@ -680,8 +680,7 @@ int MultiDevSSAGraphBuilder::CreateDistTrainOp(ir::Graph *result,
   }
 
   if (node->Op()->Type() == "split_byref" ||
-      node->Op()->Type() == "split_selected_rows" ||
-      node->Op()->Type() == "split_ids") {
+      node->Op()->Type() == "split_selected_rows") {
     // TODO(paddle-dev): getting the first var is not safe.
     op_dev_id = GetVarDeviceID(*result, input_var_names[0]);
     if (strategy_.reduce_ == BuildStrategy::ReduceStrategy::kAllReduce) {
diff --git a/paddle/fluid/framework/framework.proto b/paddle/fluid/framework/framework.proto
index c99406799ba5f664c4b9f80e0567b293e4ffea51..efdabffb9b33ddf007c13008d0f3afb7a3961eda 100644
--- a/paddle/fluid/framework/framework.proto
+++ b/paddle/fluid/framework/framework.proto
@@ -35,6 +35,7 @@ enum AttrType {
   BLOCK = 8;
   LONG = 9;
   BLOCKS = 10;
+  LONGS = 11;
 }
 
 // OpDesc describes an instance of a C++ framework::OperatorBase
@@ -55,6 +56,7 @@ message OpDesc {
     optional int32 block_idx = 12;
     optional int64 l = 13;
     repeated int32 blocks_idx = 14;
+    repeated int64 longs = 15;
   };
 
   message Var {
diff --git a/paddle/fluid/framework/op_desc.cc b/paddle/fluid/framework/op_desc.cc
index c293cf92b4f3d530109c76850df184af9cad7399..8ece618f3f72552fedcffab3e03ebb30476b7cab 100644
--- a/paddle/fluid/framework/op_desc.cc
+++ b/paddle/fluid/framework/op_desc.cc
@@ -419,8 +419,15 @@ struct SetAttrDescVisitor : public boost::static_visitor<void> {
     }
     VectorToRepeated(blocks_idx, attr_->mutable_blocks_idx());
   }
+
   void operator()(BlockDesc *desc) const { attr_->set_block_idx(desc->ID()); }
+
   void operator()(int64_t v) const { attr_->set_l(v); }
+
+  void operator()(const std::vector<int64_t> &v) const {
+    VectorToRepeated(v, attr_->mutable_longs());
+  }
+
   void operator()(boost::blank) const { PADDLE_THROW("Unexpected branch"); }
 };
 
diff --git a/paddle/fluid/framework/type_defs.h b/paddle/fluid/framework/type_defs.h
index e099e40f121ff13657e563eb608feecbca0551be..2de6233a9e0d320ec9a06d547db3575eb61925c0 100644
--- a/paddle/fluid/framework/type_defs.h
+++ b/paddle/fluid/framework/type_defs.h
@@ -36,7 +36,7 @@ using Attribute =
     boost::variant<boost::blank, int, float, std::string, std::vector<int>,
                    std::vector<float>, std::vector<std::string>, bool,
                    std::vector<bool>, BlockDesc*, int64_t,
-                   std::vector<BlockDesc*>>;
+                   std::vector<BlockDesc*>, std::vector<int64_t>>;
 
 using AttributeMap = std::unordered_map<std::string, Attribute>;
 
diff --git a/paddle/fluid/operators/fill_constant_op.cc b/paddle/fluid/operators/fill_constant_op.cc
index e04a68717b351ddb0be5a7e70aa9297e5eb0125f..252f313440296bd9e5eebf26f67b08bbe7decce8 100644
--- a/paddle/fluid/operators/fill_constant_op.cc
+++ b/paddle/fluid/operators/fill_constant_op.cc
@@ -24,7 +24,7 @@ class FillConstantInferShape : public framework::InferShapeBase {
   void operator()(framework::InferShapeContext *ctx) const override {
     PADDLE_ENFORCE(ctx->HasOutput("Out"),
                    "Output(Out) of FillConstantOp should not be null.");
-    auto &shape = ctx->Attrs().Get<std::vector<int>>("shape");
+    auto &shape = ctx->Attrs().Get<std::vector<int64_t>>("shape");
     ctx->SetOutputDim("Out", framework::make_ddim(shape));
   }
 };
@@ -47,10 +47,10 @@ class FillConstantOp : public framework::OperatorBase {
 
     if (out_var.IsType<framework::LoDTensor>()) {
       tensor = out_var.GetMutable<framework::LoDTensor>();
-      tensor->Resize(framework::make_ddim(Attr<std::vector<int>>("shape")));
+      tensor->Resize(framework::make_ddim(Attr<std::vector<int64_t>>("shape")));
     } else if (out_var.IsType<framework::SelectedRows>()) {
       tensor = out_var.GetMutable<framework::SelectedRows>()->mutable_value();
-      tensor->Resize(framework::make_ddim(Attr<std::vector<int>>("shape")));
+      tensor->Resize(framework::make_ddim(Attr<std::vector<int64_t>>("shape")));
     } else {
       PADDLE_THROW(
           "fill constant op's output only"
@@ -83,7 +83,8 @@ class FillConstantOpMaker : public framework::OpProtoAndCheckerMaker {
                  "(int, default 5 (FP32)) "
                  "Output data type")
         .SetDefault(framework::proto::VarType::FP32);
-    AddAttr<std::vector<int>>("shape", "(vector<int>) The shape of the output");
+    AddAttr<std::vector<int64_t>>("shape",
+                                  "(vector<int64_t>) The shape of the output");
     AddAttr<float>("value", "(float, default 0) The value to be filled")
         .SetDefault(0.0f);
     AddAttr<bool>("force_cpu",
diff --git a/paddle/fluid/operators/gaussian_random_op.cc b/paddle/fluid/operators/gaussian_random_op.cc
index 1488aab1926b5b4ba7bceed582700f5a11fc6c93..c70d5b8bc7569c38cbc003aca7c62dc503df11cf 100644
--- a/paddle/fluid/operators/gaussian_random_op.cc
+++ b/paddle/fluid/operators/gaussian_random_op.cc
@@ -52,7 +52,7 @@ class GaussianRandomOp : public framework::OperatorWithKernel {
   void InferShape(framework::InferShapeContext* ctx) const override {
     PADDLE_ENFORCE(ctx->HasOutput("Out"),
                    "Output(Out) of GaussianRandomOp should not be null.");
-    auto shape = ctx->Attrs().Get<std::vector<int>>("shape");
+    auto shape = ctx->Attrs().Get<std::vector<int64_t>>("shape");
     std::vector<int64_t> temp;
     temp.reserve(shape.size());
     for (auto dim : shape) {
@@ -88,9 +88,9 @@ class GaussianRandomOpMaker : public framework::OpProtoAndCheckerMaker {
   void Make() override {
     AddOutput("Out", "Output matrix of gaussian random op");
 
-    AddAttr<std::vector<int>>("shape",
-                              "(vector<int>) "
-                              "The dimension of random tensor.");
+    AddAttr<std::vector<int64_t>>("shape",
+                                  "(vector<int64_t>) "
+                                  "The dimension of random tensor.");
     AddAttr<float>("mean",
                    "(float, default 0.0) "
                    "mean of random tensor.")
diff --git a/paddle/fluid/operators/split_selected_rows_op.cc b/paddle/fluid/operators/split_selected_rows_op.cc
index 76615a9405d7a8e3fa9dba8d01a956209e02ae8f..0e7b1463d1ba81aed53e0e3f3a90d2a1fbf0ffbc 100644
--- a/paddle/fluid/operators/split_selected_rows_op.cc
+++ b/paddle/fluid/operators/split_selected_rows_op.cc
@@ -22,9 +22,9 @@ class SplitSelectedRowsOpMaker : public framework::OpProtoAndCheckerMaker {
   void Make() override {
     AddInput("X", "The input SelectedRows.");
     AddOutput("Out", "The outputs of the input SelectedRows.").AsDuplicable();
-    AddAttr<std::vector<int>>("height_sections",
-                              "Height for each output SelectedRows.")
-        .SetDefault(std::vector<int>({}));
+    AddAttr<std::vector<int64_t>>("height_sections",
+                                  "Height for each output SelectedRows.")
+        .SetDefault(std::vector<int64_t>({}));
 
     AddComment(R"DOC(
 Split a SelectedRows with a specified rows section.
diff --git a/paddle/fluid/operators/split_selected_rows_op.h b/paddle/fluid/operators/split_selected_rows_op.h
index 0e9ce165b98845f4745ee70b028513ea31cc6657..af64607fafc6544047714e731846a2440be219b8 100644
--- a/paddle/fluid/operators/split_selected_rows_op.h
+++ b/paddle/fluid/operators/split_selected_rows_op.h
@@ -21,7 +21,7 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
-static int FindOutIdx(int row, const std::vector<int>& abs_sections) {
+static int FindOutIdx(int row, const std::vector<int64_t>& abs_sections) {
   for (size_t i = 1; i < abs_sections.size(); ++i) {
     if (row < abs_sections[i]) {
       return i - 1;
@@ -30,9 +30,9 @@ static int FindOutIdx(int row, const std::vector<int>& abs_sections) {
   return abs_sections.size() - 1;
 }
 
-static std::vector<int> ToAbsoluteSection(
-    const std::vector<int>& height_sections) {
-  std::vector<int> abs_sections;
+static std::vector<int64_t> ToAbsoluteSection(
+    const std::vector<int64_t>& height_sections) {
+  std::vector<int64_t> abs_sections;
   abs_sections.resize(height_sections.size());
   abs_sections[0] = 0;
   for (size_t i = 1; i < height_sections.size(); ++i) {
@@ -47,7 +47,7 @@ class SplitSelectedRowsOpKernel : public framework::OpKernel<T> {
   void Compute(const framework::ExecutionContext& ctx) const override {
     auto* x = ctx.Input<framework::SelectedRows>("X");
     auto outs = ctx.MultiOutput<framework::SelectedRows>("Out");
-    auto height_sections = ctx.Attr<std::vector<int>>("height_sections");
+    auto height_sections = ctx.Attr<std::vector<int64_t>>("height_sections");
 
     auto abs_sections = ToAbsoluteSection(height_sections);
 
diff --git a/paddle/fluid/operators/uniform_random_op.cc b/paddle/fluid/operators/uniform_random_op.cc
index aa907595cb7cf165974caa69fe8eb0370471732d..e3132ae76f624f3338d749e4fcebbd0ecd7ffe79 100644
--- a/paddle/fluid/operators/uniform_random_op.cc
+++ b/paddle/fluid/operators/uniform_random_op.cc
@@ -29,7 +29,7 @@ class CPUUniformRandomKernel : public framework::OpKernel<T> {
     if (out_var->IsType<framework::LoDTensor>()) {
       tensor = out_var->GetMutable<framework::LoDTensor>();
     } else if (out_var->IsType<framework::SelectedRows>()) {
-      auto shape = ctx.Attr<std::vector<int>>("shape");
+      auto shape = ctx.Attr<std::vector<int64_t>>("shape");
       auto *selected_rows = out_var->GetMutable<framework::SelectedRows>();
       tensor = selected_rows->mutable_value();
       tensor->Resize(framework::make_ddim(shape));
@@ -67,7 +67,7 @@ class UniformRandomOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE(
         ctx->Attrs().Get<float>("min") < ctx->Attrs().Get<float>("max"),
         "uniform_random's min must less then max");
-    auto &shape = ctx->Attrs().Get<std::vector<int>>("shape");
+    auto &shape = ctx->Attrs().Get<std::vector<int64_t>>("shape");
     std::vector<int64_t> temp;
     temp.reserve(shape.size());
     for (auto dim : shape) {
@@ -94,7 +94,7 @@ This operator initializes a tensor with random values sampled from a
 uniform distribution. The random result is in set [min, max].
 
 )DOC");
-    AddAttr<std::vector<int>>("shape", "The shape of the output tensor");
+    AddAttr<std::vector<int64_t>>("shape", "The shape of the output tensor");
     AddAttr<float>("min", "Minimum value of uniform random. [default -1.0].")
         .SetDefault(-1.0f);
     AddAttr<float>("max", "Maximun value of uniform random. [default 1.0].")
diff --git a/paddle/fluid/operators/uniform_random_op.cu b/paddle/fluid/operators/uniform_random_op.cu
index bbb692b0ddfc18e8a62c0d2a6bac88f9932f6704..2bb0ecc139f7096d1b61150e0a2d4fb095338749 100644
--- a/paddle/fluid/operators/uniform_random_op.cu
+++ b/paddle/fluid/operators/uniform_random_op.cu
@@ -48,7 +48,7 @@ class GPUUniformRandomKernel : public framework::OpKernel<T> {
     if (out_var->IsType<framework::LoDTensor>()) {
       tensor = out_var->GetMutable<framework::LoDTensor>();
     } else if (out_var->IsType<framework::SelectedRows>()) {
-      auto shape = context.Attr<std::vector<int>>("shape");
+      auto shape = context.Attr<std::vector<int64_t>>("shape");
       tensor = out_var->GetMutable<framework::SelectedRows>()->mutable_value();
       tensor->Resize(framework::make_ddim(shape));
     } else {
diff --git a/paddle/fluid/pybind/protobuf.cc b/paddle/fluid/pybind/protobuf.cc
index 3b22718a8c6f994dbc2dc3e7aaa19a7163f716ba..cbc83106fc36d3e03a05d1172b4574c09178435b 100644
--- a/paddle/fluid/pybind/protobuf.cc
+++ b/paddle/fluid/pybind/protobuf.cc
@@ -259,6 +259,8 @@ void BindOpDesc(pybind11::module *m) {
   pybind11::enum_<pd::proto::AttrType>(*m, "AttrType", "")
       .value("INT", pd::proto::AttrType::INT)
       .value("INTS", pd::proto::AttrType::INTS)
+      .value("LONG", pd::proto::AttrType::LONG)
+      .value("LONGS", pd::proto::AttrType::LONGS)
       .value("FLOAT", pd::proto::AttrType::FLOAT)
       .value("FLOATS", pd::proto::AttrType::FLOATS)
       .value("STRING", pd::proto::AttrType::STRING)