diff --git a/paddle/fluid/framework/dlpack_tensor.cc b/paddle/fluid/framework/dlpack_tensor.cc
index 057a19f31759b47f61b7a17a71920183f06ec18d..b7bca733b8f9e63bd5bc88ecced6b55c2078e886 100644
--- a/paddle/fluid/framework/dlpack_tensor.cc
+++ b/paddle/fluid/framework/dlpack_tensor.cc
@@ -69,7 +69,8 @@ static DLDataType GetDLDataTypeFromTypeIndex(proto::VarType::Type type) {
 #undef REG_DL_DATA_TYPE
 }
 
-struct DLDeviceVisitor : public boost::static_visitor<::DLDevice> {
+struct DLDeviceVisitor
+    : public std::unary_function<const platform::Place &, ::DLDevice> {
   inline ::DLDevice operator()(const platform::CPUPlace &place) const {
     ::DLDevice device;
     device.device_type = kDLCPU;
diff --git a/paddle/fluid/framework/ir/generate_pass.cc b/paddle/fluid/framework/ir/generate_pass.cc
index 160304784a9fa69d3248f773cf48a90470ffd3f2..455af83427819afaf90c1d7f387e4856be7930a6 100644
--- a/paddle/fluid/framework/ir/generate_pass.cc
+++ b/paddle/fluid/framework/ir/generate_pass.cc
@@ -21,7 +21,7 @@ namespace paddle {
 namespace framework {
 namespace ir {
 
-class element_visitor : public boost::static_visitor<Attribute> {
+class element_visitor {
  public:
   explicit element_visitor(int index) : index_(index) {}
 
@@ -47,7 +47,7 @@ class element_visitor : public boost::static_visitor<Attribute> {
   int index_;
 };
 
-class operation_visitor : public boost::static_visitor<Attribute> {
+class operation_visitor {
  public:
   explicit operation_visitor(const proto::PassDesc::OperationType& type)
       : type_(type) {}
diff --git a/paddle/fluid/framework/op_desc.cc b/paddle/fluid/framework/op_desc.cc
index 169722b971b29fc97b41040b69810dc9c7bb0e37..c0a9528c28126ff3f8abc50066ca28640f5354e5 100644
--- a/paddle/fluid/framework/op_desc.cc
+++ b/paddle/fluid/framework/op_desc.cc
@@ -764,7 +764,7 @@ void OpDesc::RenameInput(const std::string &old_name,
   need_update_ = true;
 }
 
-struct SetAttrDescVisitor : public boost::static_visitor<void> {
+struct SetAttrDescVisitor {
   explicit SetAttrDescVisitor(proto::OpDesc::Attr *attr) : attr_(attr) {}
   mutable proto::OpDesc::Attr *attr_;
   void operator()(int v) const { attr_->set_i(v); }
diff --git a/paddle/fluid/framework/tensor_util.cc b/paddle/fluid/framework/tensor_util.cc
index dd80458b624c6c8857515252f638fc7b1725788d..dbb549efa2519d48136648ba0cfbba2c4d82cc9a 100644
--- a/paddle/fluid/framework/tensor_util.cc
+++ b/paddle/fluid/framework/tensor_util.cc
@@ -689,7 +689,7 @@ inline void AnyImpl(Predicate predicate,
 }
 
 template <typename Predicate>
-class AnyVisitor : public boost::static_visitor<bool> {
+class AnyVisitor : public std::unary_function<const Place&, bool> {
  private:
   const framework::Tensor& tensor_;
   Predicate predicate_;
@@ -774,7 +774,7 @@ class AnyVisitor : public boost::static_visitor<bool> {
 };
 
 template <typename Predicate>
-class AnyOutVisitor : public boost::static_visitor<> {
+class AnyOutVisitor : public std::unary_function<const Place&, void> {
  private:
   const framework::Tensor& tensor_;
   mutable framework::Tensor* out_;
@@ -843,7 +843,7 @@ inline void AllImpl(Predicate predicate,
 }
 
 template <typename Predicate>
-class AllOutVisitor : public boost::static_visitor<> {
+class AllOutVisitor : public std::unary_function<const Place&, void> {
  private:
   const framework::Tensor& tensor_;
   mutable framework::Tensor* out_;
@@ -942,7 +942,7 @@ static inline void __global__ BothFalse(const T* cmp, T* out, int element_num) {
 }
 #endif
 
-struct BothFalseVisitor : public boost::static_visitor<> {
+struct BothFalseVisitor : public std::unary_function<const Place&, void> {
   const framework::Tensor& in_;
   mutable framework::Tensor* out_;
   BothFalseVisitor(const framework::Tensor& in, framework::Tensor* out)
diff --git a/paddle/fluid/imperative/gradient_accumulator.cc b/paddle/fluid/imperative/gradient_accumulator.cc
index 4a8fc6a5d546ce150f1b46de7885e1a9f250dd91..f6883fe6c6a923429235874ab788e7bd2f224c1f 100644
--- a/paddle/fluid/imperative/gradient_accumulator.cc
+++ b/paddle/fluid/imperative/gradient_accumulator.cc
@@ -79,7 +79,8 @@ static void MoveOrCopyVar(framework::Variable* dst,
 }
 
 template <typename T>
-class TensorAddFunctor : public boost::static_visitor<> {
+class TensorAddFunctor
+    : public std::unary_function<const platform::Place&, void> {
  public:
   TensorAddFunctor(int64_t numel, const T* x, T* y)
       : numel_(numel), x_(x), y_(y) {}
diff --git a/paddle/fluid/memory/allocation/naive_best_fit_allocator.cc b/paddle/fluid/memory/allocation/naive_best_fit_allocator.cc
index 9d5f048a1651db0b16c05cafca79afddbcb9a1a0..57c5941d5227d9943568ec4ac8f387bfb1666889 100644
--- a/paddle/fluid/memory/allocation/naive_best_fit_allocator.cc
+++ b/paddle/fluid/memory/allocation/naive_best_fit_allocator.cc
@@ -59,7 +59,7 @@ uint64_t Release(const Place &place);
 template <typename Place>
 size_t Used(const Place &place);
 
-struct Usage : public boost::static_visitor<size_t> {
+struct Usage {
   size_t operator()(const platform::CPUPlace &cpu) const;
   size_t operator()(const platform::CUDAPlace &gpu) const;
   size_t operator()(const platform::CUDAPinnedPlace &cuda_pinned) const;
@@ -894,7 +894,7 @@ size_t Used<platform::CustomPlace>(const platform::CustomPlace &place) {
 #endif
 }
 
-struct AllocVisitor : public boost::static_visitor<void *> {
+struct AllocVisitor : std::unary_function<const Place, void *> {
   inline explicit AllocVisitor(size_t size) : size_(size) {}
 
   template <typename Place>
@@ -906,7 +906,7 @@ struct AllocVisitor : public boost::static_visitor<void *> {
   size_t size_;
 };
 
-struct FreeVisitor : public boost::static_visitor<void> {
+struct FreeVisitor : public std::unary_function<const Place, void> {
   inline explicit FreeVisitor(void *ptr, size_t size)
       : ptr_(ptr), size_(size) {}
 
@@ -920,7 +920,7 @@ struct FreeVisitor : public boost::static_visitor<void> {
   size_t size_;
 };
 
-struct ReleaseVisitor : public boost::static_visitor<uint64_t> {
+struct ReleaseVisitor : std::unary_function<const Place, uint64_t> {
   template <typename Place>
   inline uint64_t operator()(const Place &place) const {
     return Release<Place>(place);
diff --git a/paddle/fluid/operators/array_to_lod_tensor_op.cc b/paddle/fluid/operators/array_to_lod_tensor_op.cc
index 5b23ff604759acb302144c6506e6404e371051e6..a2af64e2276801a05c114c7b7c6477949d781df8 100644
--- a/paddle/fluid/operators/array_to_lod_tensor_op.cc
+++ b/paddle/fluid/operators/array_to_lod_tensor_op.cc
@@ -43,7 +43,7 @@ struct ArrayToLoDFunctorImpl {
   void apply();
 };
 
-struct ArrayToLoDFunctor : public boost::static_visitor<void> {
+struct ArrayToLoDFunctor : public std::unary_function<platform::Place, void> {
   std::vector<framework::Tensor> in;
   mutable framework::Tensor *out;
 
diff --git a/paddle/fluid/operators/controlflow/feed_op.cc b/paddle/fluid/operators/controlflow/feed_op.cc
index 00806d18c066fee1106cec10b8426c0b8b8eb28b..4cef104496510f4c75d2857cadc1538eb10476cc 100644
--- a/paddle/fluid/operators/controlflow/feed_op.cc
+++ b/paddle/fluid/operators/controlflow/feed_op.cc
@@ -29,7 +29,7 @@ namespace operators {
 
 // FeedVariableVisitor is to feed the variable data
 // according to data type (LoDTensor or  Strings).
-class FeedVariableVisitor : public boost::static_visitor<void> {
+class FeedVariableVisitor {
  public:
   explicit FeedVariableVisitor(framework::Variable *out_var,
                                const platform::Place &place)
diff --git a/paddle/fluid/operators/controlflow/op_variant.cc b/paddle/fluid/operators/controlflow/op_variant.cc
index 60f58955adbedaa0a01fb7e698e173bd48bf2068..48b7a4341067284a4c48e2c9e89c26e077d49570 100644
--- a/paddle/fluid/operators/controlflow/op_variant.cc
+++ b/paddle/fluid/operators/controlflow/op_variant.cc
@@ -17,24 +17,21 @@
 namespace paddle {
 namespace operators {
 
-struct InputsVisitor
-    : public boost::static_visitor<const framework::VariableNameMap *> {
+struct InputsVisitor {
   template <typename OpType>
   const framework::VariableNameMap *operator()(const OpType *op) const {
     return &(op->Inputs());
   }
 };
 
-struct OutputsVisitor
-    : public boost::static_visitor<const framework::VariableNameMap *> {
+struct OutputsVisitor {
   template <typename OpType>
   const framework::VariableNameMap *operator()(const OpType *op) const {
     return &(op->Outputs());
   }
 };
 
-struct AttributeMapVisitor
-    : public boost::static_visitor<const framework::AttributeMap *> {
+struct AttributeMapVisitor {
   const framework::AttributeMap *operator()(const framework::OpDesc *op) const {
     return &(op->GetAttrMap());
   }
@@ -45,7 +42,7 @@ struct AttributeMapVisitor
   }
 };
 
-struct RawPointerVisitor : public boost::static_visitor<const void *> {
+struct RawPointerVisitor {
   template <typename OpType>
   const void *operator()(const OpType *op) const {
     return op;
diff --git a/paddle/fluid/operators/lod_tensor_to_array_op.cc b/paddle/fluid/operators/lod_tensor_to_array_op.cc
index 147b23f56acdcfca0c3daf265a3121a6fbbe42e7..d4b36f31e6201f8e52eb388481eb80a3dfcbfa46 100644
--- a/paddle/fluid/operators/lod_tensor_to_array_op.cc
+++ b/paddle/fluid/operators/lod_tensor_to_array_op.cc
@@ -44,7 +44,8 @@ struct LoDTensorToArrayFunctorImpl {
   void apply();
 };
 
-struct LoDTensorToArrayFunctor : public boost::static_visitor<void> {
+struct LoDTensorToArrayFunctor
+    : public std::unary_function<platform::Place, void> {
   std::vector<const framework::Tensor *> ref_inputs_;
   mutable std::vector<framework::Tensor *> outputs_;
   const framework::Tensor &input_;
diff --git a/paddle/fluid/operators/math/matrix_bit_code.cc b/paddle/fluid/operators/math/matrix_bit_code.cc
index 8a6f098baefd9539425fa4882a95a34b72c92728..0648f2497d9d75c4771014ec285718263c3f96cb 100644
--- a/paddle/fluid/operators/math/matrix_bit_code.cc
+++ b/paddle/fluid/operators/math/matrix_bit_code.cc
@@ -19,7 +19,7 @@ namespace operators {
 namespace math {
 
 template <typename T>
-struct MatrixBitCodeFunctorAdd : public boost::static_visitor<void> {
+struct MatrixBitCodeFunctorAdd {
   const framework::Tensor &vec_;
   framework::Tensor *tmat_;
 
@@ -51,7 +51,7 @@ void MatrixBitCodeFunctor<T>::Add(const framework::Tensor &vec,
 }
 
 template <typename T>
-struct MatrixBitCodeFunctorAddGrad : public boost::static_visitor<void> {
+struct MatrixBitCodeFunctorAddGrad {
   const framework::Tensor &tmat_;
   framework::Tensor *vec_;
   MatrixBitCodeFunctorAddGrad(const framework::Tensor &tmat,
@@ -83,7 +83,7 @@ void MatrixBitCodeFunctor<T>::AddGrad(const framework::Tensor &tmat,
 }
 
 template <typename T>
-struct MatrixBitCodeFunctorSum : public boost::static_visitor<void> {
+struct MatrixBitCodeFunctorSum {
   const framework::Tensor &tmat_;
   framework::Tensor *sum_;
   T scale_sum_;
@@ -125,7 +125,7 @@ void MatrixBitCodeFunctor<T>::Sum(const framework::Tensor &tmat,
 }
 
 template <typename T>
-struct MatrixBitCodeFunctorMul : public boost::static_visitor<void> {
+struct MatrixBitCodeFunctorMul {
   framework::Tensor *tmat_;
   const framework::Tensor &weight_;
   const framework::Tensor &input_;
@@ -174,7 +174,7 @@ class ReservedVector : public std::vector<T> {
 };
 
 template <typename T>
-struct MatrixBitCodeFunctorMulGradWeight : public boost::static_visitor<void> {
+struct MatrixBitCodeFunctorMulGradWeight {
   const framework::Tensor &tmat_;
   framework::Tensor *weight_;
   const framework::Tensor &input_;
@@ -224,8 +224,7 @@ void MatrixBitCodeFunctor<T>::MulGradWeight(const framework::Tensor &tmat,
 }
 
 template <typename T>
-struct MatrixBitCodeFunctorMulGradWeightSR
-    : public boost::static_visitor<void> {
+struct MatrixBitCodeFunctorMulGradWeightSR {
   const framework::Tensor &tmat_;
   phi::SelectedRows *weight_;
   const framework::Tensor &input_;
@@ -280,7 +279,7 @@ void MatrixBitCodeFunctor<T>::MulGradWeight(const framework::Tensor &tmat,
 }
 
 template <typename T>
-struct MatrixBitCodeFunctorMulGradError : public boost::static_visitor<void> {
+struct MatrixBitCodeFunctorMulGradError {
   const framework::Tensor &tmat_;
   const framework::Tensor &weight_;
   framework::Tensor *input_;
@@ -324,7 +323,7 @@ void MatrixBitCodeFunctor<T>::MulGradError(const framework::Tensor &tmat,
 }
 
 template <typename T>
-struct MatrixBitCodeFunctorSub : public boost::static_visitor<void> {
+struct MatrixBitCodeFunctorSub {
   framework::Tensor *tmat_;
 
   explicit MatrixBitCodeFunctorSub(framework::Tensor *tmat) : tmat_(tmat) {}
diff --git a/paddle/fluid/platform/device/ipu/ipu_compiler.cc b/paddle/fluid/platform/device/ipu/ipu_compiler.cc
index 39ff4601b6749e021e5776dc7ef4982fb8e75542..330ddef577ef20f85805311d2640fe5ada1d0145 100644
--- a/paddle/fluid/platform/device/ipu/ipu_compiler.cc
+++ b/paddle/fluid/platform/device/ipu/ipu_compiler.cc
@@ -32,7 +32,7 @@ namespace ipu {
 
 namespace {
 
-struct CustomOpAttrVisitor : public boost::static_visitor<void> {
+struct CustomOpAttrVisitor {
   CustomOpAttrVisitor(std::map<std::string, popart::any>* attr,
                       const std::string& attr_name)
       : attrs_(attr), attr_name_(attr_name) {}
@@ -82,7 +82,7 @@ struct CustomOpAttrVisitor : public boost::static_visitor<void> {
   }
 };
 
-struct ConstantOpAttrVisitor : public boost::static_visitor<void> {
+struct ConstantOpAttrVisitor {
   ConstantOpAttrVisitor(framework::LoDTensor* tensor, VarType::Type dtype)
       : tensor_(tensor), dtype_(dtype) {}
 
diff --git a/paddle/fluid/platform/variant.h b/paddle/fluid/platform/variant.h
index fb4772abd30621087bf45acecce78de476b3e85e..9682749898fc77af54d81c1dccdac3b474dcd9ca 100644
--- a/paddle/fluid/platform/variant.h
+++ b/paddle/fluid/platform/variant.h
@@ -40,7 +40,6 @@ limitations under the License. */
 
 #include <boost/mpl/comparison.hpp>
 #include <boost/mpl/less_equal.hpp>
-#include <boost/variant.hpp>
 
 #include "paddle/utils/any.h"
 #include "paddle/utils/optional.h"
diff --git a/paddle/fluid/pybind/global_value_getter_setter.cc b/paddle/fluid/pybind/global_value_getter_setter.cc
index 2871d1de56780f70777f43fa730cbdc34843ec77..b2a52e568aed9fa317b0592ac0b3f8f6b71ab5af 100644
--- a/paddle/fluid/pybind/global_value_getter_setter.cc
+++ b/paddle/fluid/pybind/global_value_getter_setter.cc
@@ -217,7 +217,7 @@ void BindGlobalValueGetterSetter(pybind11::module *module) {
                        GlobalVarGetterSetterRegistry::CreateSetter(&var)); \
   } while (0)
 
-struct RegisterGetterSetterVisitor : public boost::static_visitor<void> {
+struct RegisterGetterSetterVisitor {
   RegisterGetterSetterVisitor(const std::string &name,
                               bool is_writable,
                               void *value_ptr)
diff --git a/paddle/fluid/pybind/pybind_boost_headers.h b/paddle/fluid/pybind/pybind_boost_headers.h
index 623ec84acda6f67409b50c565fa0abea1fac8d17..2a25990944d141c3be962b6f901219ed0b177cba 100644
--- a/paddle/fluid/pybind/pybind_boost_headers.h
+++ b/paddle/fluid/pybind/pybind_boost_headers.h
@@ -18,14 +18,12 @@ limitations under the License. */
 #include <vector>
 
 #include "glog/logging.h"
-#include "paddle/fluid/platform/variant.h"
 #include "paddle/utils/variant.h"
 #include "pybind11/numpy.h"
 #include "pybind11/pybind11.h"
 #include "pybind11/stl.h"
 // Cast paddle::variant for PyBind.
 // Copy from
-
 // https://github.com/pybind/pybind11/issues/576#issuecomment-269563199
 namespace pybind11 {
 namespace detail {
@@ -78,10 +76,7 @@ struct paddle_variant_caster<V<Ts...>> {
   using Type = V<Ts...>;
 
   template <typename T>
-  typename std::enable_if<
-      !std::is_same<T, boost::detail::variant::void_>::value,
-      bool>::type
-  try_load(handle src, bool convert) {
+  bool try_load(handle src, bool convert) {
     auto caster = make_caster<T>();
     if (!load_success_ && caster.load(src, convert)) {
       load_success_ = true;
@@ -112,13 +107,6 @@ struct paddle_variant_caster<V<Ts...>> {
     return false;
   }
 
-  template <typename T>
-  typename std::enable_if<std::is_same<T, boost::detail::variant::void_>::value,
-                          bool>::type
-  try_load(handle src, bool convert) {
-    return false;
-  }
-
   bool load(handle src, bool convert) {
     auto unused = {false, try_load<Ts>(src, convert)...};
     (void)(unused);
@@ -128,11 +116,6 @@ struct paddle_variant_caster<V<Ts...>> {
   static handle cast(Type const& src,
                      return_value_policy policy,
                      handle parent) {
-    /*
-    auto paddle_variant_caster_visitor = [&](Type const& src)->handle {
-      return make_caster<Type>::cast(src, policy, parent);
-    }
-    */
     paddle_variant_caster_visitor visitor(policy, parent);
     return paddle::visit(visitor, src);
   }
diff --git a/paddle/phi/kernels/funcs/math_function.cc b/paddle/phi/kernels/funcs/math_function.cc
index 61cedb57faf2721b7332065272251ba6e1e18ffd..15a708f02f4974c0d8a5e3e485276c789ff37f41 100644
--- a/paddle/phi/kernels/funcs/math_function.cc
+++ b/paddle/phi/kernels/funcs/math_function.cc
@@ -214,7 +214,8 @@ void set_constant_with_place<paddle::platform::CUDAPinnedPlace>(
   phi::VisitDataType(tensor->dtype(), TensorSetConstantCPU(tensor, value));
 }
 
-struct TensorSetConstantWithPlace : public boost::static_visitor<void> {
+struct TensorSetConstantWithPlace
+    : public std::unary_function<paddle::platform::Place, void> {
   TensorSetConstantWithPlace(const paddle::platform::DeviceContext& context,
                              paddle::framework::Tensor* tensor,
                              float value)