diff --git a/paddle/fluid/framework/custom_kernel_test.cc b/paddle/fluid/framework/custom_kernel_test.cc
index 708b7bbe8a5295085c25f8796d97f1e5cddc717a..5f01681624ec65d431337ed729f274ea76a8f169 100644
--- a/paddle/fluid/framework/custom_kernel_test.cc
+++ b/paddle/fluid/framework/custom_kernel_test.cc
@@ -212,11 +212,13 @@ TEST(CustomKernel, custom_kernel_dot) {
   kernel_context.EmplaceBackAttr(fake_attr_int64_vec);
   kernel_context.EmplaceBackAttr(fake_attr_int_vec);
 
-  auto out_meta = pten::DotInferMeta(dense_x->meta(), dense_y->meta());
   auto dense_out = std::make_shared<pten::DenseTensor>(
       pten::make_intrusive<paddle::experimental::SharedStorage>(
           pten::TransToFluidPlace(backend)),
-      std::move(out_meta));
+      pten::DenseTensorMeta());
+
+  pten::MetaTensor meta_out(dense_out.get());
+  pten::DotInferMeta(*dense_x, *dense_y, &meta_out);
   kernel_context.EmplaceBackOutput(dense_out.get());  // idx:0 index:[0,1)
 
   // fake_input_vec: idx:1, index:[1,3)
diff --git a/paddle/fluid/framework/infershape_utils.cc b/paddle/fluid/framework/infershape_utils.cc
index 80787c3b87121da0508cac7f3e0069e066ed534e..e1d7190a9e418ab5ae049707e2d2d92d78a896be 100644
--- a/paddle/fluid/framework/infershape_utils.cc
+++ b/paddle/fluid/framework/infershape_utils.cc
@@ -186,6 +186,14 @@ class CompatMetaTensor : public pten::MetaTensor {
     }
   }
 
+  void share_meta(const MetaTensor& meta_tensor) override {
+    set_dims(meta_tensor.dims());
+    set_dtype(meta_tensor.dtype());
+    // VarDesc doesn't contains layout, so we cannot share layout
+    // set_layout(meta_tensor.layout());
+    share_lod(meta_tensor);
+  }
+
  private:
   const LoD& GetRuntimeLoD() const {
     auto* var = BOOST_GET_CONST(Variable*, var_);
diff --git a/paddle/pten/api/lib/api_utils.h b/paddle/pten/api/lib/api_utils.h
index 6332132fac2fd0c529eeef6754d173d33195d39d..3d18cc611cf60db1161cfc367b9539197c94d5bf 100644
--- a/paddle/pten/api/lib/api_utils.h
+++ b/paddle/pten/api/lib/api_utils.h
@@ -18,6 +18,7 @@ limitations under the License. */
 #include "paddle/pten/api/lib/utils/storage.h"
 #include "paddle/pten/core/compat/convert_utils.h"
 #include "paddle/pten/core/dense_tensor.h"
+#include "paddle/pten/core/meta_tensor.h"
 
 namespace paddle {
 namespace experimental {
@@ -44,44 +45,38 @@ inline std::unique_ptr<std::vector<pten::DenseTensor>> TensorToDenseTensor(
 
 /* ----------------- for infer_meta --------------------- */
 
-inline const pten::DenseTensorMeta& GetDenseTensorMeta(
-    const pten::DenseTensor& tensor) {
-  return tensor.meta();
+inline pten::MetaTensor MakeMetaTensor(const pten::DenseTensor& tensor) {
+  return pten::MetaTensor(tensor);
 }
 
-inline std::vector<pten::DenseTensorMeta> GetDenseTensorMeta(
+inline std::vector<pten::MetaTensor> MakeMetaTensor(
     const std::vector<pten::DenseTensor>& tensors) {
-  std::vector<pten::DenseTensorMeta> metas;
-  metas.reserve(tensors.size());
+  std::vector<pten::MetaTensor> meta_tensors;
+  meta_tensors.reserve(tensors.size());
   for (const auto& t : tensors) {
-    metas.push_back(t.meta());
+    meta_tensors.emplace_back(t);
   }
-  return metas;
+  return meta_tensors;
 }
 
 /* ------------------ for output ----------------------- */
 
-inline pten::DenseTensor* SetKernelOutput(const pten::DenseTensorMeta& meta,
-                                          Backend backend,
-                                          Tensor* out) {
+inline pten::DenseTensor* SetKernelOutput(Backend backend, Tensor* out) {
   auto dense_tensor = std::make_shared<pten::DenseTensor>(
       pten::make_intrusive<SharedStorage>(pten::TransToFluidPlace(backend)),
-      meta);
+      pten::DenseTensorMeta());
   out->set_impl(dense_tensor);
   return dense_tensor.get();
 }
 
 inline std::vector<pten::DenseTensor*> SetKernelOutput(
-    const std::vector<pten::DenseTensorMeta>& metas,
-    Backend backend,
-    std::vector<Tensor>* out) {
-  size_t n = metas.size();
-  out->reserve(n);
-  std::vector<pten::DenseTensor*> results(n);
-  for (size_t i = 0; i < n; ++i) {
+    size_t out_size, Backend backend, std::vector<Tensor>* out) {
+  out->reserve(out_size);
+  std::vector<pten::DenseTensor*> results(out_size);
+  for (size_t i = 0; i < out_size; ++i) {
     auto tensor_ptr = std::make_shared<pten::DenseTensor>(
         pten::make_intrusive<SharedStorage>(pten::TransToFluidPlace(backend)),
-        metas[i]);
+        pten::DenseTensorMeta());
     results[i] = tensor_ptr.get();
     out->emplace_back();
     out->back().set_impl(tensor_ptr);
diff --git a/paddle/pten/api/lib/manual_api.cc b/paddle/pten/api/lib/manual_api.cc
index c7d05533bbf977e9e6ae980ddf2e38745d408d16..5b697c3ff763d3277edf391c945f9f7293fd0373 100644
--- a/paddle/pten/api/lib/manual_api.cc
+++ b/paddle/pten/api/lib/manual_api.cc
@@ -57,20 +57,19 @@ PADDLE_API Tensor copy_to(const Tensor& x, Backend backend, bool blocking) {
   kernel_context.EmplaceBackInput(dense_x.get());
   kernel_context.EmplaceBackAttr(blocking);
 
-  // 4. InferMeta
-  auto out_meta = UnchangedInferMeta(dense_x->meta());
-
-  // 5. Prepare outputs
+  // 4. Prepare outputs & InferMeta
   auto dense_out = std::make_shared<pten::DenseTensor>(
       pten::make_intrusive<paddle::experimental::SharedStorage>(
           pten::TransToFluidPlace(backend)),
-      std::move(out_meta));
+      pten::DenseTensorMeta());
+  pten::MetaTensor meta_out(dense_out.get());
+  pten::UnchangedInferMeta(*dense_x, &meta_out);
   dense_out->mutable_data(pten::TransToFluidPlace(backend));
   kernel_context.EmplaceBackOutput(dense_out.get());
   Tensor out;
   out.set_impl(dense_out);
 
-  // 6. Call kernel
+  // 5. Call kernel
   kernel(&kernel_context);
 
   return out;
diff --git a/paddle/pten/core/infermeta_utils.h b/paddle/pten/core/infermeta_utils.h
index 0f45185c8340d2001e624db4bdbb5758e02ff6a1..fecfab7153f53178aed4458e4239d5331fb6f74e 100644
--- a/paddle/pten/core/infermeta_utils.h
+++ b/paddle/pten/core/infermeta_utils.h
@@ -26,16 +26,6 @@ limitations under the License. */
 
 namespace pten {
 
-// TODO(chenweihang): add other flags if needed
-struct MetaConfig {
-  bool is_runtime{true};
-
-  MetaConfig() = default;
-
-  // supporting implicit construction is easier to use
-  MetaConfig(bool is_runtime) : is_runtime(is_runtime) {}  // NOLINT
-};
-
 class InferMetaContext {
  public:
   InferMetaContext() = default;
diff --git a/paddle/pten/core/meta_tensor.cc b/paddle/pten/core/meta_tensor.cc
index fd558d20a064cef595ef673aca10f95ca20d6534..d205ee1ca40b0d88d9393d1c5ba5415474fe0298 100644
--- a/paddle/pten/core/meta_tensor.cc
+++ b/paddle/pten/core/meta_tensor.cc
@@ -33,7 +33,7 @@ void MetaTensor::set_dims(const DDim& dims) {
     DenseTensorUtils::GetMutableMeta(static_cast<DenseTensor*>(tensor_))->dims =
         dims;
   } else {
-    PADDLE_THROW(paddle::platform::errors::Unimplemented(
+    PADDLE_THROW(pten::errors::Unimplemented(
         "Unsupported setting dims for `%s`.", tensor_->type_info().name()));
   }
 }
@@ -43,7 +43,7 @@ void MetaTensor::set_dtype(DataType dtype) {
     DenseTensorUtils::GetMutableMeta(static_cast<DenseTensor*>(tensor_))
         ->dtype = dtype;
   } else {
-    PADDLE_THROW(paddle::platform::errors::Unimplemented(
+    PADDLE_THROW(pten::errors::Unimplemented(
         "Unsupported settting dtype for `%s`.", tensor_->type_info().name()));
   }
 }
@@ -53,7 +53,7 @@ void MetaTensor::set_layout(DataLayout layout) {
     DenseTensorUtils::GetMutableMeta(static_cast<DenseTensor*>(tensor_))
         ->layout = layout;
   } else {
-    PADDLE_THROW(paddle::platform::errors::Unimplemented(
+    PADDLE_THROW(pten::errors::Unimplemented(
         "Unsupported settting layout for `%s`.", tensor_->type_info().name()));
   }
 }
@@ -63,9 +63,9 @@ void MetaTensor::share_lod(const MetaTensor& meta_tensor) {
     DenseTensorUtils::GetMutableMeta(static_cast<DenseTensor*>(tensor_))->lod =
         meta_tensor.lod();
   } else {
-    PADDLE_THROW(paddle::platform::errors::Unimplemented(
-        "Unsupported share lod inplace for `%s`.",
-        tensor_->type_info().name()));
+    PADDLE_THROW(
+        pten::errors::Unimplemented("Unsupported sharing lod inplace for `%s`.",
+                                    tensor_->type_info().name()));
   }
 }
 
@@ -73,8 +73,20 @@ const LoD& MetaTensor::lod() const {
   if (pten::DenseTensor::classof(tensor_)) {
     return static_cast<DenseTensor*>(tensor_)->lod();
   } else {
-    PADDLE_THROW(paddle::platform::errors::Unimplemented(
-        "Unsupported setting dims for `%s`.", tensor_->type_info().name()));
+    PADDLE_THROW(pten::errors::Unimplemented("Unsupported getting lod of `%s`.",
+                                             tensor_->type_info().name()));
+  }
+}
+
+void MetaTensor::share_meta(const MetaTensor& meta_tensor) {
+  if (pten::DenseTensor::classof(tensor_)) {
+    set_dims(meta_tensor.dims());
+    set_dtype(meta_tensor.dtype());
+    set_layout(meta_tensor.layout());
+    share_lod(meta_tensor);
+  } else {
+    PADDLE_THROW(pten::errors::Unimplemented(
+        "Unsupported sharing meta for `%s`.", tensor_->type_info().name()));
   }
 }
 
diff --git a/paddle/pten/core/meta_tensor.h b/paddle/pten/core/meta_tensor.h
index ae2ec606267dc45e18e6b6a102b560d2b91adf2d..6ccb698fe18e4f6c4936aa94c487496b4c60b013 100644
--- a/paddle/pten/core/meta_tensor.h
+++ b/paddle/pten/core/meta_tensor.h
@@ -23,11 +23,26 @@ limitations under the License. */
 
 namespace pten {
 
+// TODO(chenweihang): add other flags if needed
+struct MetaConfig {
+  bool is_runtime{true};
+
+  MetaConfig() = default;
+
+  // supporting implicit construction is easier to use
+  MetaConfig(bool is_runtime) : is_runtime(is_runtime) {}  // NOLINT
+};
+
 class MetaTensor {
  public:
-  explicit MetaTensor(TensorBase* tensor) : tensor_(tensor) {}
-
   MetaTensor() = default;
+
+  // supporting implicit construction is easier to use
+  MetaTensor(TensorBase* tensor) : tensor_(tensor) {}  // NOLINT
+  MetaTensor(const TensorBase& tensor)                 // NOLINT
+      : tensor_(const_cast<TensorBase*>(&tensor)) {}
+  MetaTensor(TensorBase& tensor) : tensor_(&tensor) {}  // NOLINT
+
   MetaTensor(const MetaTensor&) = default;
   MetaTensor(MetaTensor&&) = default;
   MetaTensor& operator=(const MetaTensor&) = delete;
@@ -42,7 +57,9 @@ class MetaTensor {
   virtual void set_dims(const DDim& dims);
   virtual void set_dtype(DataType dtype);
   virtual void set_layout(DataLayout layout);
+
   virtual void share_lod(const MetaTensor& meta_tensor);
+  virtual void share_meta(const MetaTensor& meta_tensor);
 
  private:
   // Because the lod in compiletime and runtime is different,
diff --git a/paddle/pten/infermeta/CMakeLists.txt b/paddle/pten/infermeta/CMakeLists.txt
index 2216d38708b0b4746e55481ca63299b96b496eb6..c077e7b4c55636e07eaf9353d009e857c239b8ec 100644
--- a/paddle/pten/infermeta/CMakeLists.txt
+++ b/paddle/pten/infermeta/CMakeLists.txt
@@ -1,2 +1,2 @@
-cc_library(infermeta SRCS nullary.cc unary.cc binary.cc multiary.cc DEPS convert_utils infermeta_utils)
-cc_library(backward_infermeta SRCS backward.cc DEPS convert_utils)
+cc_library(infermeta SRCS nullary.cc unary.cc binary.cc multiary.cc DEPS convert_utils meta_tensor infermeta_utils)
+cc_library(backward_infermeta SRCS backward.cc DEPS meta_tensor convert_utils)
diff --git a/paddle/pten/infermeta/backward.cc b/paddle/pten/infermeta/backward.cc
index 5a66e8cd2ecb384be7bd807269cce290e8f8e04e..b7bb17bdd1c38b1d2616bf31c52ecfdbd8626b55 100644
--- a/paddle/pten/infermeta/backward.cc
+++ b/paddle/pten/infermeta/backward.cc
@@ -16,13 +16,15 @@ limitations under the License. */
 
 namespace pten {
 
-std::tuple<DenseTensorMeta, DenseTensorMeta> MatmulGradInferMeta(
-    const DenseTensorMeta& x_meta,
-    const DenseTensorMeta& y_meta,
-    const DenseTensorMeta& out_grad_meta,
-    bool transpose_x,
-    bool transpose_y) {
-  return std::make_tuple(x_meta, y_meta);
+void MatmulGradInferMeta(const MetaTensor& x,
+                         const MetaTensor& y,
+                         const MetaTensor& out_grad_meta,
+                         bool transpose_x,
+                         bool transpose_y,
+                         MetaTensor* dx,
+                         MetaTensor* dy) {
+  dx->share_meta(x);
+  dy->share_meta(y);
 }
 
 }  // namespace pten
diff --git a/paddle/pten/infermeta/backward.h b/paddle/pten/infermeta/backward.h
index 03bdb3a962a96e84f7ed569c18d3b73fad145a78..d6b96861412861de6fb892a28c3930bd7db20da7 100644
--- a/paddle/pten/infermeta/backward.h
+++ b/paddle/pten/infermeta/backward.h
@@ -15,15 +15,17 @@ limitations under the License. */
 #pragma once
 
 #include <tuple>
-#include "paddle/pten/core/tensor_meta.h"
+
+#include "paddle/pten/core/meta_tensor.h"
 
 namespace pten {
 
-std::tuple<DenseTensorMeta, DenseTensorMeta> MatmulGradInferMeta(
-    const DenseTensorMeta& x_meta,
-    const DenseTensorMeta& y_meta,
-    const DenseTensorMeta& out_grad_meta,
-    bool transpose_x,
-    bool transpose_y);
+void MatmulGradInferMeta(const MetaTensor& x,
+                         const MetaTensor& y,
+                         const MetaTensor& out_grad_meta,
+                         bool transpose_x,
+                         bool transpose_y,
+                         MetaTensor* dx,
+                         MetaTensor* dy);
 
 }  // namespace pten
diff --git a/paddle/pten/infermeta/binary.cc b/paddle/pten/infermeta/binary.cc
index cb605db78d962e2deff1295686c5e95945f02531..02d78b5caa722923f34f34e98357654ddad11ba0 100644
--- a/paddle/pten/infermeta/binary.cc
+++ b/paddle/pten/infermeta/binary.cc
@@ -12,15 +12,13 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-// See Note [ Why still include the fluid headers? ]
 #include "paddle/pten/infermeta/binary.h"
 #include "paddle/pten/kernels/funcs/common_shape.h"
 
 namespace pten {
 
-DenseTensorMeta DotInferMeta(const DenseTensorMeta& x_meta,
-                             const DenseTensorMeta& y_meta) {
-  auto x_dims = x_meta.dims;
+void DotInferMeta(const MetaTensor& x, const MetaTensor& y, MetaTensor* out) {
+  auto x_dims = x.dims();
   auto x_rank = static_cast<size_t>(x_dims.size());
   PADDLE_ENFORCE_EQ(true,
                     1 == x_rank || 2 == x_rank,
@@ -29,10 +27,10 @@ DenseTensorMeta DotInferMeta(const DenseTensorMeta& x_meta,
                         "should be 1 or 2",
                         x_dims.to_str()));
 
-  auto y_dims = y_meta.dims;
+  auto y_dims = y.dims();
   PADDLE_ENFORCE_EQ(
       true,
-      x_rank == (size_t)y_dims.size(),
+      x_rank == static_cast<size_t>(y_dims.size()),
       paddle::platform::errors::PreconditionNotMet(
           "ShapeError: The shape of input tensor Y: %s should match with "
           "input tenosr X: %s",
@@ -56,25 +54,27 @@ DenseTensorMeta DotInferMeta(const DenseTensorMeta& x_meta,
                         y_dims.to_str()));
 
   x_dims[x_dims.size() - 1] = 1;
-  DenseTensorMeta return_meta(x_meta.dtype, x_dims, x_meta.layout);
-  return return_meta;
+  out->set_dims(x_dims);
+  out->set_dtype(x.dtype());
+  out->set_layout(x.layout());
 }
 
-DenseTensorMeta MatmulInferMeta(const DenseTensorMeta& x_meta,
-                                const DenseTensorMeta& y_meta,
-                                bool trans_x,
-                                bool trans_y) {
-  std::vector<int64_t> dims_x = pten::framework::vectorize(x_meta.dims);
-  std::vector<int64_t> dims_y = pten::framework::vectorize(y_meta.dims);
+void MatmulInferMeta(const MetaTensor& x,
+                     const MetaTensor& y,
+                     bool trans_x,
+                     bool trans_y,
+                     MetaTensor* out) {
+  std::vector<int64_t> dims_x = pten::framework::vectorize(x.dims());
+  std::vector<int64_t> dims_y = pten::framework::vectorize(y.dims());
   auto ndims_x = dims_x.size();
   auto ndims_y = dims_y.size();
   PADDLE_ENFORCE_GT(ndims_x,
-                    0,
+                    0UL,
                     paddle::platform::errors::InvalidArgument(
                         "The Input(x) dims size must be greater than 0,"
                         " but reviced dims size is 0. "));
   PADDLE_ENFORCE_GT(ndims_y,
-                    0,
+                    0UL,
                     paddle::platform::errors::InvalidArgument(
                         "The Input(y) dims size must be greater than 0,"
                         " but reviced dims size is 0. "));
@@ -127,21 +127,24 @@ DenseTensorMeta MatmulInferMeta(const DenseTensorMeta& x_meta,
 
   auto ddim_out = pten::framework::make_ddim(new_dims);
 
-  return {x_meta.dtype, ddim_out, x_meta.layout};
+  out->set_dims(ddim_out);
+  out->set_dtype(x.dtype());
+  out->set_layout(x.layout());
 }
 
-DenseTensorMeta ElementwiseInferMeta(const DenseTensorMeta& x_meta,
-                                     const DenseTensorMeta& y_meta) {
-  return ElementwiseRawInferMeta(x_meta, y_meta, -1);
+void ElementwiseInferMeta(const MetaTensor& x,
+                          const MetaTensor& y,
+                          MetaTensor* out) {
+  return ElementwiseRawInferMeta(x, y, -1, std::move(out));
 }
 
-DenseTensorMeta ElementwiseRawInferMeta(const DenseTensorMeta& x_meta,
-                                        const DenseTensorMeta& y_meta,
-                                        int axis) {
-  DenseTensorMeta return_meta(x_meta.dtype, x_meta.dims, x_meta.layout);
-  if (x_meta.dims != y_meta.dims) {
-    auto x_dims = x_meta.dims;
-    auto y_dims = y_meta.dims;
+void ElementwiseRawInferMeta(const MetaTensor& x,
+                             const MetaTensor& y,
+                             int axis,
+                             MetaTensor* out) {
+  if (x.dims() != y.dims()) {
+    auto x_dims = x.dims();
+    auto y_dims = y.dims();
     int max_dim = std::max(x_dims.size(), y_dims.size());
     if (x_dims.size() == y_dims.size()) {
       PADDLE_ENFORCE_EQ((axis == -1) || (axis == 0),
@@ -174,10 +177,15 @@ DenseTensorMeta ElementwiseRawInferMeta(const DenseTensorMeta& x_meta,
                                   out_dims_array.data(),
                                   max_dim,
                                   axis);
-    return_meta.dims = pten::framework::make_ddim(out_dims_array);
+    auto out_dims = pten::framework::make_ddim(out_dims_array);
+    out->set_dims(out_dims);
+  } else {
+    out->set_dims(x.dims());
   }
-  return_meta.lod = x_meta.lod;
-  return return_meta;
+
+  out->set_dtype(x.dtype());
+  out->set_layout(x.layout());
+  out->share_lod(x);
 }
 
 }  // namespace pten
diff --git a/paddle/pten/infermeta/binary.h b/paddle/pten/infermeta/binary.h
index 658211e48ac0a44c57d83ce63154e481a90ce69c..992082464967c74461f77ba4616470f0a09fc15b 100644
--- a/paddle/pten/infermeta/binary.h
+++ b/paddle/pten/infermeta/binary.h
@@ -14,38 +14,35 @@ limitations under the License. */
 
 #pragma once
 
-// See Note [ Why still include the fluid headers? ]
-#include "paddle/pten/core/tensor_meta.h"
+#include "paddle/pten/core/meta_tensor.h"
 
 namespace pten {
 
 // Common InferMeta Functions for binary operators, The format like:
 //
-//   1. DenseTensorMeta [OpName]InferMeta(const DenseTensorMeta& x_meta, ...)
-//   {}
-//   2. std::pair<DenseTensorMeta, DenseTensorMeta> [OpName]InferMeta(const
-//   DenseTensorMeta&
-//   x_meta, ...) {}
-//   3. std::tuple<DenseTensorMeta, DenseTensorMeta, DenseTensorMeta>
-//   [OpName]InferMeta(const
-//   DenseTensorMeta& x_meta, ...)
-//  NOTE: The name "InferMeta" may be not appropriate. "InferMeta" may be good.
-//  Because functions in this file
-//  not only can infer shape, but alse need infer lod or other useful data.
-
-DenseTensorMeta DotInferMeta(const DenseTensorMeta& x_meta,
-                             const DenseTensorMeta& y_meta);
-
-DenseTensorMeta MatmulInferMeta(const DenseTensorMeta& x_meta,
-                                const DenseTensorMeta& y_meta,
-                                bool trans_x,
-                                bool trans_y);
-
-DenseTensorMeta ElementwiseInferMeta(const DenseTensorMeta& x_meta,
-                                     const DenseTensorMeta& y_meta);
-
-DenseTensorMeta ElementwiseRawInferMeta(const DenseTensorMeta& x_meta,
-                                        const DenseTensorMeta& y_meta,
-                                        int axis);
-
+//   1. void [FunctionDesc|OpName]InferMeta(const MetaTensor& x,
+//                                          const MetaTensor& y,
+//                                          ...,
+//                                          MetaTensor* out) {}
+//
+// NOTE: The name "InferShape" may be not appropriate. "InferMeta" may be good.
+//   Because functions in this file not only can infer shape, but also need
+//   infer lod or other useful data.
+
+void DotInferMeta(const MetaTensor& x, const MetaTensor& y, MetaTensor* out);
+
+void MatmulInferMeta(const MetaTensor& x,
+                     const MetaTensor& y,
+                     bool trans_x,
+                     bool trans_y,
+                     MetaTensor* out);
+
+void ElementwiseInferMeta(const MetaTensor& x,
+                          const MetaTensor& y,
+                          MetaTensor* out);
+
+void ElementwiseRawInferMeta(const MetaTensor& x_meta,
+                             const MetaTensor& y_meta,
+                             int axis,
+                             MetaTensor* out);
 }  // namespace pten
diff --git a/paddle/pten/infermeta/multiary.cc b/paddle/pten/infermeta/multiary.cc
index ecd0396a28688f5473e9ba0144e53a944ec135d7..869e87df5d8de4c53e51e726e3f6f4b6bf80c829 100644
--- a/paddle/pten/infermeta/multiary.cc
+++ b/paddle/pten/infermeta/multiary.cc
@@ -18,18 +18,19 @@ limitations under the License. */
 #include "paddle/pten/kernels/funcs/concat_funcs.h"
 namespace pten {
 
-DenseTensorMeta ConcatInferMeta(const std::vector<DenseTensorMeta>& x_meta,
-                                const Scalar& axis_scalar,
-                                bool is_runtime) {
-  PADDLE_ENFORCE_GE(x_meta.size(),
-                    0,
+void ConcatInferMeta(const std::vector<MetaTensor>& x,
+                     const Scalar& axis_scalar,
+                     MetaTensor* out,
+                     MetaConfig config) {
+  PADDLE_ENFORCE_GE(x.size(),
+                    0UL,
                     paddle::platform::errors::InvalidArgument(
                         "The size of input meta vector should be greater"
                         "than 0."));
 
   int axis = axis_scalar.to<int>();
   // 1. calculate axis
-  int rank = x_meta[0].dims.size();
+  int rank = x.at(0).dims().size();
   PADDLE_ENFORCE_EQ(
       axis >= -rank && axis < rank,
       true,
@@ -44,13 +45,15 @@ DenseTensorMeta ConcatInferMeta(const std::vector<DenseTensorMeta>& x_meta,
 
   // 2. calculate out dims
   std::vector<pten::DDim> x_dims;
-  for (auto meta : x_meta) {
-    x_dims.push_back(meta.dims);
+  for (auto& x_t : x) {
+    x_dims.push_back(x_t.dims());
   }
   pten::DDim out_dim =
-      pten::funcs::ComputeAndCheckShape(is_runtime, x_dims, axis);
+      pten::funcs::ComputeAndCheckShape(config.is_runtime, x_dims, axis);
 
-  return {x_meta[0].dtype, out_dim, x_meta[0].layout};
+  out->set_dims(out_dim);
+  out->set_dtype(x.at(0).dtype());
+  out->set_layout(x.at(0).layout());
 }
 
 }  // namespace pten
diff --git a/paddle/pten/infermeta/multiary.h b/paddle/pten/infermeta/multiary.h
index f8d5468e50d47f787bdaba4deaecf8c53503926f..c251699da4a94c75555886dfebf803497706e1c5 100644
--- a/paddle/pten/infermeta/multiary.h
+++ b/paddle/pten/infermeta/multiary.h
@@ -15,12 +15,12 @@ limitations under the License. */
 #pragma once
 
 #include "paddle/pten/common/scalar.h"
-#include "paddle/pten/core/tensor_meta.h"
+#include "paddle/pten/core/meta_tensor.h"
 namespace pten {
 
-// TODO(chentianyu03) use std::vector<DenseTensor> as InferMeta inputs
-DenseTensorMeta ConcatInferMeta(const std::vector<DenseTensorMeta>& x_meta,
-                                const Scalar& axis_scalar,
-                                bool is_runtime);
+void ConcatInferMeta(const std::vector<MetaTensor>& x,
+                     const Scalar& axis_scalar,
+                     MetaTensor* out,
+                     MetaConfig config = MetaConfig());
 
 }  // namespace pten
diff --git a/paddle/pten/infermeta/nullary.cc b/paddle/pten/infermeta/nullary.cc
index 19e11f049fee78bab3a30b678bf7ecd67cff4847..fd9b2a8f717f47aec0a80e0ede68508520153cf7 100644
--- a/paddle/pten/infermeta/nullary.cc
+++ b/paddle/pten/infermeta/nullary.cc
@@ -12,23 +12,25 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-// See Note [ Why still include the fluid headers? ]
 #include "paddle/pten/infermeta/nullary.h"
 
 namespace pten {
 
-DenseTensorMeta CreateInferMeta(const std::vector<int64_t>& shape,
-                                DataType dtype,
-                                DataLayout layout) {
-  const auto& out_dims = pten::framework::make_ddim(shape);
-  return {dtype, out_dims, layout};
+void CreateInferMeta(const std::vector<int64_t>& shape,
+                     DataType dtype,
+                     DataLayout layout,
+                     MetaTensor* out) {
+  auto out_dims = pten::framework::make_ddim(shape);
+  out->set_dims(out_dims);
+  out->set_dtype(dtype);
+  out->set_layout(layout);
 }
 
-DenseTensorMeta CreateInferMeta(const ScalarArray& shape,
-                                DataType dtype,
-                                DataLayout layout) {
-  const auto& out_dims = pten::framework::make_ddim(shape.GetData());
-  return {dtype, out_dims, layout};
+void CreateInferMeta(const ScalarArray& shape,
+                     DataType dtype,
+                     DataLayout layout,
+                     MetaTensor* out) {
+  CreateInferMeta(shape.GetData(), dtype, layout, out);
 }
 
 }  // namespace pten
diff --git a/paddle/pten/infermeta/nullary.h b/paddle/pten/infermeta/nullary.h
index 721a39bb3ac31a972345903b091e890eebc71c53..f0b6aad26bea6a0ff606f1d353ca4649042234e7 100644
--- a/paddle/pten/infermeta/nullary.h
+++ b/paddle/pten/infermeta/nullary.h
@@ -15,24 +15,27 @@ limitations under the License. */
 #pragma once
 
 #include "paddle/pten/common/scalar_array.h"
-#include "paddle/pten/core/tensor_meta.h"
+#include "paddle/pten/core/meta_tensor.h"
 
 namespace pten {
 
 // Common InferMeta Functions for 0-nary operators(no input tensor), The format
 // like:
 //
-//   1. DenseTensorMeta [OpName]InferMeta( ...)
-//  NOTE: The name "InferMeta" may be not appropriate. "InferMeta" may be good.
-//  Because functions in this file
-//  not only can infer shape, but alse need infer lod or other useful data.
-
-DenseTensorMeta CreateInferMeta(const std::vector<int64_t>& shape,
-                                DataType dtype,
-                                DataLayout layout);
-
-DenseTensorMeta CreateInferMeta(const ScalarArray& shape,
-                                DataType dtype,
-                                DataLayout layout);
+//   1. void [FunctionDesc|OpName]InferMeta(..., MetaTensor* out)
+//
+// NOTE: The name "InferShape" may be not appropriate. "InferMeta" may be good.
+//   Because functions in this file not only can infer shape, but also need
+//   infer lod or other useful data.
+
+void CreateInferMeta(const std::vector<int64_t>& shape,
+                     DataType dtype,
+                     DataLayout layout,
+                     MetaTensor* out);
+
+void CreateInferMeta(const ScalarArray& shape,
+                     DataType dtype,
+                     DataLayout layout,
+                     MetaTensor* out);
 
 }  // namespace pten
diff --git a/paddle/pten/infermeta/unary.cc b/paddle/pten/infermeta/unary.cc
index e1a30d10e160be62e96d5ef70cb64528c8daf043..57bac52cef59199cfd4e21ab18ecd99bb3c4a7e3 100644
--- a/paddle/pten/infermeta/unary.cc
+++ b/paddle/pten/infermeta/unary.cc
@@ -16,6 +16,7 @@ limitations under the License. */
 
 #include <set>
 
+#include "paddle/pten/common/data_type.h"
 #include "paddle/pten/core/infermeta_utils.h"
 
 namespace pten {
@@ -23,26 +24,22 @@ namespace pten {
 void UnchangedInferMetaNew(MetaConfig config,
                            const MetaTensor& x,
                            MetaTensor* out) {
-  out->set_dims(x.dims());
-  out->set_dtype(x.dtype());
-  out->set_layout(x.layout());
-  out->share_lod(x);
+  out->share_meta(x);
 }
 
 DenseTensorMeta UnchangedInferMeta(const DenseTensorMeta& x_meta) {
   return x_meta;
 }
 
-DenseTensorMeta ReductionInferMeta(const DenseTensorMeta& x_meta) {
-  const auto& out_dims = pten::framework::make_ddim({1});
-  DenseTensorMeta return_meta(x_meta.dtype, out_dims, x_meta.layout);
-  return return_meta;
+void UnchangedInferMeta(const MetaTensor& x, MetaTensor* out) {
+  out->share_meta(x);
 }
 
-DenseTensorMeta FlattenInferMeta(const DenseTensorMeta& x_meta,
-                                 int start_axis,
-                                 int stop_axis) {
-  auto& x_dims = x_meta.dims;
+void FlattenInferMeta(const MetaTensor& x,
+                      int start_axis,
+                      int stop_axis,
+                      MetaTensor* out) {
+  auto x_dims = x.dims();
   int in_dims_size = x_dims.size();
   if (start_axis < 0) {
     start_axis = start_axis + in_dims_size;
@@ -75,29 +72,30 @@ DenseTensorMeta FlattenInferMeta(const DenseTensorMeta& x_meta,
     out_shape.push_back(x_dims[i]);
   }
   const auto& out_dims = pten::framework::make_ddim(out_shape);
-  DenseTensorMeta return_meta(x_meta.dtype, out_dims, x_meta.layout);
+  out->set_dims(out_dims);
+  out->set_dtype(x.dtype());
+  out->set_layout(x.layout());
 
-  if (x_dims[0] == return_meta.dims[0]) {
+  if (x_dims[0] == out_dims[0]) {
     // Only pass LoD when the first dimension of output and Input(X)
     // are the same.
-    return_meta.lod = x_meta.lod;
+    out->share_lod(x);
   }
-
-  return return_meta;
 }
 
-DenseTensorMeta CastInferMeta(const DenseTensorMeta& x_meta,
-                              const DataType out_dtype) {
-  DenseTensorMeta out_meta(out_dtype, x_meta.dims, x_meta.layout);
-  return out_meta;
+void CastInferMeta(const MetaTensor& x, DataType out_dtype, MetaTensor* out) {
+  out->set_dims(x.dims());
+  out->set_dtype(out_dtype);
+  out->set_layout(x.layout());
 }
 
-DenseTensorMeta CreateLikeInferMeta(const DenseTensorMeta& x_meta,
-                                    DataType dtype,
-                                    DataLayout layout) {
-  return {dtype == DataType::UNDEFINED ? x_meta.dtype : dtype,
-          x_meta.dims,
-          layout == DataLayout::UNDEFINED ? x_meta.layout : layout};
+void CreateLikeInferMeta(const MetaTensor& x,
+                         DataType dtype,
+                         DataLayout layout,
+                         MetaTensor* out) {
+  out->set_dims(x.dims());
+  out->set_dtype(dtype == DataType::UNDEFINED ? x.dtype() : dtype);
+  out->set_layout(layout == DataLayout::UNDEFINED ? x.layout() : layout);
 }
 
 static pten::framework::DDim ValidateShape(
@@ -220,46 +218,51 @@ static pten::framework::DDim ValidateShape(
   return pten::framework::make_ddim(output_shape);
 }
 
-DenseTensorMeta InferMetaFromVecValue(const DenseTensorMeta& x_meta,
-                                      const std::vector<int64_t>& shape) {
+void InferMetaFromVecValue(const MetaTensor& x,
+                           const std::vector<int64_t>& shape,
+                           MetaTensor* out) {
   PADDLE_ENFORCE_EQ(!shape.empty(),
                     true,
                     paddle::platform::errors::InvalidArgument(
                         "The parameter 'shape' in ReshapeOp must be set. "
                         "But received 'shape' is empty."));
-  auto x_dims = x_meta.dims;
+  auto x_dims = x.dims();
   auto out_dims = ValidateShape(shape, x_dims);
-  DenseTensorMeta return_meta(x_meta.dtype, out_dims, x_meta.layout);
-  if (x_dims[0] == return_meta.dims[0]) {
+  out->set_dims(out_dims);
+  out->set_dtype(x.dtype());
+  out->set_layout(x.layout());
+  if (x_dims[0] == out_dims[0]) {
     // Only pass LoD when the first dimension of output and Input(X)
     // are the same.
-    return_meta.lod = x_meta.lod;
+    out->share_lod(x);
   }
-  return return_meta;
 }
 
-DenseTensorMeta ReshapeInferMeta(const DenseTensorMeta& x_meta,
-                                 const ScalarArray& shape) {
-  return InferMetaFromVecValue(x_meta, shape.GetData());
+void ReshapeInferMeta(const MetaTensor& x,
+                      const ScalarArray& shape,
+                      MetaTensor* out) {
+  InferMetaFromVecValue(x, shape.GetData(), out);
 }
 
 /*  Why not use ReduceInferMeta directly?
     Because we need make InferMetaFunction's args follow the design of api.yaml
 */
-DenseTensorMeta SumInferMeta(const DenseTensorMeta& x_meta,
-                             const std::vector<int64_t>& axis,
-                             DataType dtype,
-                             bool keep_dim) {
-  return ReduceInferMeta(x_meta, axis, keep_dim, dtype);
+void SumInferMeta(const MetaTensor& x,
+                  const std::vector<int64_t>& axis,
+                  DataType dtype,
+                  bool keep_dim,
+                  MetaTensor* out) {
+  ReduceInferMeta(x, axis, keep_dim, dtype, std::move(out));
 }
 
-DenseTensorMeta ReduceInferMeta(const DenseTensorMeta& x_meta,
-                                const std::vector<int64_t>& axis,
-                                bool keep_dim,
-                                DataType dtype) {
+void ReduceInferMeta(const MetaTensor& x,
+                     const std::vector<int64_t>& axis,
+                     bool keep_dim,
+                     DataType dtype,
+                     MetaTensor* out) {
   bool reduce_all = true;
   std::set<int64_t> dims_set(axis.begin(), axis.end());
-  for (int64_t i = 0; i < x_meta.dims.size(); ++i) {
+  for (int64_t i = 0; i < x.dims().size(); ++i) {
     if (dims_set.find(i) == dims_set.end()) {
       reduce_all = false;
       break;
@@ -268,19 +271,19 @@ DenseTensorMeta ReduceInferMeta(const DenseTensorMeta& x_meta,
 
   std::vector<int64_t> out_dim_vector;
   if (keep_dim) {
-    for (int64_t i = 0; i < x_meta.dims.size(); ++i) {
+    for (int64_t i = 0; i < x.dims().size(); ++i) {
       if (reduce_all || dims_set.find(i) != dims_set.end()) {
         out_dim_vector.push_back(1);
       } else {
-        out_dim_vector.push_back(x_meta.dims.at(i));
+        out_dim_vector.push_back(x.dims().at(i));
       }
     }
   } else {
-    for (int64_t i = 0; i < x_meta.dims.size(); ++i) {
+    for (int64_t i = 0; i < x.dims().size(); ++i) {
       if (reduce_all || dims_set.find(i) != dims_set.end()) {
         continue;
       } else {
-        out_dim_vector.push_back(x_meta.dims.at(i));
+        out_dim_vector.push_back(x.dims().at(i));
       }
     }
 
@@ -294,16 +297,24 @@ DenseTensorMeta ReduceInferMeta(const DenseTensorMeta& x_meta,
   if (dtype != DataType::UNDEFINED) {
     out_dtype = dtype;
   } else {
-    if (x_meta.dtype == DataType::BOOL || x_meta.dtype == DataType::INT32 ||
-        x_meta.dtype == DataType::INT64) {
+    if (x.dtype() == DataType::BOOL || x.dtype() == DataType::INT32 ||
+        x.dtype() == DataType::INT64) {
       out_dtype = DataType::INT64;
     } else {
-      out_dtype = x_meta.dtype;
+      out_dtype = x.dtype();
     }
   }
 
-  DenseTensorMeta return_meta(out_dtype, out_dim, x_meta.layout);
-  return return_meta;
+  out->set_dims(out_dim);
+  out->set_dtype(out_dtype);
+  out->set_layout(x.layout());
+}
+
+void ReduceInferMeta(const MetaTensor& x,
+                     const std::vector<int64_t>& axis,
+                     bool keep_dim,
+                     MetaTensor* out) {
+  ReduceInferMeta(x, axis, keep_dim, DataType::UNDEFINED, out);
 }
 
 }  // namespace pten
diff --git a/paddle/pten/infermeta/unary.h b/paddle/pten/infermeta/unary.h
index 670c70de84ccfdd01288ad5ad02b0b0ce5226c24..c1a939c2dec503246a7f8acec512578a4b9b225f 100644
--- a/paddle/pten/infermeta/unary.h
+++ b/paddle/pten/infermeta/unary.h
@@ -16,9 +16,7 @@ limitations under the License. */
 
 // See Note [ Why still include the fluid headers? ]
 #include "paddle/pten/common/scalar_array.h"
-#include "paddle/pten/core/infermeta_utils.h"
 #include "paddle/pten/core/meta_tensor.h"
-#include "paddle/pten/core/tensor_meta.h"
 
 namespace pten {
 
@@ -26,45 +24,54 @@ class MetaConfig;
 
 // Common InferMeta Functions for unary operators, The format like:
 //
-//   void [OpName]InferMeta(const MetaTensor& x, ..., MetaTensor* out) {}
+//   void [FunctionDesc|OpName]InferMeta(const MetaTensor& x, ..., MetaTensor*
+//   out) {}
 //
 // NOTE: The name "InferShape" may be not appropriate. "InferMeta" may be good.
 // Because functions in this file not only can infer shape, but also need
 // infer lod or other useful data.
 
-// TODO(chenweihang): update all InferMeta function format in next pr,
-// now add UnchangedInferMetaNew for test new format
+// TODO(chenweihang): to avoid conflit, remove this function in next PR
 void UnchangedInferMetaNew(MetaConfig config,
                            const MetaTensor& x,
                            MetaTensor* out);
 
-DenseTensorMeta UnchangedInferMeta(const DenseTensorMeta& x_meta);
+void UnchangedInferMeta(const MetaTensor& x, MetaTensor* out);
 
-DenseTensorMeta ReductionInferMeta(const DenseTensorMeta& x_meta);
+void FlattenInferMeta(const MetaTensor& x,
+                      int start_axis,
+                      int stop_axis,
+                      MetaTensor* out);
 
-DenseTensorMeta FlattenInferMeta(const DenseTensorMeta& x_meta,
-                                 int start_axis,
-                                 int stop_axis);
-DenseTensorMeta CastInferMeta(const DenseTensorMeta& x_meta,
-                              const DataType out_dtype);
+void CastInferMeta(const MetaTensor& x, DataType out_dtype, MetaTensor* out);
 
-DenseTensorMeta CreateLikeInferMeta(const DenseTensorMeta& x_meta,
-                                    DataType dtype,
-                                    DataLayout layout);
+void CreateLikeInferMeta(const MetaTensor& x,
+                         DataType dtype,
+                         DataLayout layout,
+                         MetaTensor* out);
 
-DenseTensorMeta InferMetaFromVecValue(const DenseTensorMeta& x_meta,
-                                      const std::vector<int64_t>& shape);
-
-DenseTensorMeta ReshapeInferMeta(const DenseTensorMeta& x_meta,
-                                 const ScalarArray& shape);
-
-DenseTensorMeta ReduceInferMeta(const DenseTensorMeta& x_meta,
-                                const std::vector<int64_t>& axis,
-                                bool keep_dim,
-                                DataType dtype = DataType::UNDEFINED);
+void InferMetaFromVecValue(const MetaTensor& x,
+                           const std::vector<int64_t>& shape,
+                           MetaTensor* out);
 
-DenseTensorMeta SumInferMeta(const DenseTensorMeta& x_meta,
-                             const std::vector<int64_t>& axis,
-                             DataType dtype,
-                             bool keep_dim);
+void ReshapeInferMeta(const MetaTensor& x,
+                      const ScalarArray& shape,
+                      MetaTensor* out);
+
+void ReduceInferMeta(const MetaTensor& x,
+                     const std::vector<int64_t>& axis,
+                     bool keep_dim,
+                     DataType dtype,
+                     MetaTensor* out);
+
+void ReduceInferMeta(const MetaTensor& x,
+                     const std::vector<int64_t>& axis,
+                     bool keep_dim,
+                     MetaTensor* out);
+
+void SumInferMeta(const MetaTensor& x,
+                  const std::vector<int64_t>& axis,
+                  DataType dtype,
+                  bool keep_dim,
+                  MetaTensor* out);
 }  // namespace pten
diff --git a/paddle/pten/kernels/cast_kernel.h b/paddle/pten/kernels/cast_kernel.h
index 8fdce9cda6f1d1c2f4b3ea6bb6a8c0fd592486b8..a7f84619345f2fd019ced85ef38838fa938153c4 100644
--- a/paddle/pten/kernels/cast_kernel.h
+++ b/paddle/pten/kernels/cast_kernel.h
@@ -29,8 +29,9 @@ template <typename T, typename Context>
 DenseTensor Cast(const Context& dev_ctx,
                  const DenseTensor& x,
                  DataType out_dtype) {
-  auto out_meta = CastInferMeta(x.meta(), out_dtype);
-  auto dense_out = pten::Empty<T, Context>(dev_ctx, std::move(out_meta));
+  auto dense_out = pten::Empty<T, Context>(dev_ctx);
+  MetaTensor meta_out(&dense_out);
+  CastInferMeta(x, out_dtype, &meta_out);
   CastKernel<T, Context>(dev_ctx, x, out_dtype, &dense_out);
   return dense_out;
 }
diff --git a/paddle/pten/kernels/complex_kernel.h b/paddle/pten/kernels/complex_kernel.h
index ff27144eb49d4a9db95f2ed82f644c553cefea6b..ab1cb59872a0430b122faa0bcd9816d801314bc4 100644
--- a/paddle/pten/kernels/complex_kernel.h
+++ b/paddle/pten/kernels/complex_kernel.h
@@ -32,8 +32,9 @@ template <typename T,
                   std::is_same<T, paddle::platform::complex<double>>::value,
               bool> = true>
 DenseTensor Conj(const Context& dev_ctx, const DenseTensor& x) {
-  auto out_meta = UnchangedInferMeta(x.meta());
-  auto dense_out = pten::Empty<T, Context>(dev_ctx, std::move(out_meta));
+  auto dense_out = pten::Empty<T, Context>(dev_ctx);
+  MetaTensor meta_out(&dense_out);
+  UnchangedInferMeta(x, &meta_out);
   ConjKernel<T>(dev_ctx, x, &dense_out);
   return dense_out;
 }
diff --git a/paddle/pten/kernels/concat_kernel.h b/paddle/pten/kernels/concat_kernel.h
index 310b9ba8c0c4c50521b2dcea2de62171557f83da..8c9103145f66f95b5dd5ed0a40c8512e09302a2b 100644
--- a/paddle/pten/kernels/concat_kernel.h
+++ b/paddle/pten/kernels/concat_kernel.h
@@ -30,14 +30,16 @@ template <typename T, typename Context>
 DenseTensor Concat(const Context& dev_ctx,
                    const std::vector<DenseTensor>& x,
                    const Scalar& axis) {
-  std::vector<DenseTensorMeta> x_meta;
-  for (auto t : x) {
-    x_meta.push_back(t.meta());
+  std::vector<MetaTensor> meta_x;
+  for (const auto& t : x) {
+    meta_x.emplace_back(t);
   }
 
-  auto out_meta = ConcatInferMeta(x_meta, axis.to<int>(), true);
-  auto dense_out = pten::Empty<T, Context>(dev_ctx, std::move(out_meta));
+  auto dense_out = pten::Empty<T, Context>(dev_ctx);
+  MetaTensor meta_out(&dense_out);
+  ConcatInferMeta(meta_x, axis.to<int>(), &meta_out, /*is_runtime=*/true);
   ConcatKernel<T, Context>(dev_ctx, x, axis, &dense_out);
   return dense_out;
 }
+
 }  // namespace pten
diff --git a/paddle/pten/kernels/dot_kernel.h b/paddle/pten/kernels/dot_kernel.h
index 47f1c89109e7ebd887be066f95c4de3c663c29d5..67f6ca3517fc597cce3390f7a4c220c91525f8fd 100644
--- a/paddle/pten/kernels/dot_kernel.h
+++ b/paddle/pten/kernels/dot_kernel.h
@@ -29,8 +29,9 @@ template <typename T, typename Context>
 DenseTensor Dot(const Context& dev_ctx,
                 const DenseTensor& x,
                 const DenseTensor& y) {
-  auto out_meta = DotInferMeta(x.meta(), y.meta());
-  auto dense_out = pten::Empty<T, Context>(dev_ctx, std::move(out_meta));
+  auto dense_out = pten::Empty<T, Context>(dev_ctx);
+  MetaTensor meta_out(&dense_out);
+  DotInferMeta(x, y, &meta_out);
   DotKernel<T, Context>(dev_ctx, x, y, &dense_out);
   return dense_out;
 }
diff --git a/paddle/pten/kernels/empty_kernel.h b/paddle/pten/kernels/empty_kernel.h
index d283ef5c1e41ef64a2f4a38b382595fa2db3bd90..8a7da8fbd5659fdfb5af73e03e1cce8868b9c726 100644
--- a/paddle/pten/kernels/empty_kernel.h
+++ b/paddle/pten/kernels/empty_kernel.h
@@ -55,8 +55,9 @@ DenseTensor Empty(const Context& dev_ctx,
                   DataType dtype = DataType::FLOAT32,
                   Backend backend = Backend::CPU,  // Is backend needed here?
                   DataLayout layout = DataLayout::NCHW) {
-  auto out_meta = CreateInferMeta(shape, dtype, layout);
-  auto dense_out = Empty<T, Context>(dev_ctx, std::move(out_meta));
+  auto dense_out = Empty<T, Context>(dev_ctx);
+  MetaTensor meta_out(&dense_out);
+  CreateInferMeta(shape, dtype, layout, &meta_out);
   EmptyKernel<T, Context>(dev_ctx, shape, &dense_out);
   return dense_out;
 }
@@ -68,8 +69,9 @@ DenseTensor EmptyLike(
     DataType dtype = DataType::UNDEFINED,
     Backend backend = Backend::UNDEFINED,  // Is backend needed here?
     DataLayout layout = DataLayout::UNDEFINED) {
-  auto out_meta = CreateLikeInferMeta(x.meta(), dtype, layout);
-  auto dense_out = Empty<T, Context>(dev_ctx, std::move(out_meta));
+  auto dense_out = Empty<T, Context>(dev_ctx);
+  MetaTensor meta_out(&dense_out);
+  CreateLikeInferMeta(x, dtype, layout, &meta_out);
   EmptyLikeKernel<T, Context>(dev_ctx, &dense_out);
   return dense_out;
 }
diff --git a/paddle/pten/kernels/flatten_kernel.h b/paddle/pten/kernels/flatten_kernel.h
index c974fda1ed36322c021e5695e536277def5986c3..38d8786c7fc824f5d067df042236f026d12ee3eb 100644
--- a/paddle/pten/kernels/flatten_kernel.h
+++ b/paddle/pten/kernels/flatten_kernel.h
@@ -40,8 +40,9 @@ DenseTensor Flatten(const Context& dev_ctx,
                     const DenseTensor& x,
                     int start_axis,
                     int stop_axis) {
-  auto out_meta = FlattenInferMeta(x.meta(), start_axis, stop_axis);
-  auto dense_out = Empty<T, Context>(dev_ctx, std::move(out_meta));
+  auto dense_out = Empty<T, Context>(dev_ctx);
+  MetaTensor meta_out(&dense_out);
+  FlattenInferMeta(x, start_axis, stop_axis, &meta_out);
   FlattenKernel<T, Context>(dev_ctx, x, start_axis, stop_axis, &dense_out);
   return dense_out;
 }
diff --git a/paddle/pten/kernels/full_kernel.h b/paddle/pten/kernels/full_kernel.h
index bc484fb4edffa95abf25c66172b7e6d3c603a500..030eb4b1c760f688a1c2ee11cff3664e0ef03137 100644
--- a/paddle/pten/kernels/full_kernel.h
+++ b/paddle/pten/kernels/full_kernel.h
@@ -41,8 +41,9 @@ DenseTensor Full(const Context& dev_ctx,
                  DataType dtype = DataType::FLOAT32,
                  Backend backend = Backend::CPU,  // Is backend needed here?
                  DataLayout layout = DataLayout::NCHW) {
-  auto out_meta = CreateInferMeta(shape, dtype, layout);
-  auto dense_out = Empty<T, Context>(dev_ctx, std::move(out_meta));
+  auto dense_out = Empty<T, Context>(dev_ctx);
+  MetaTensor meta_out(&dense_out);
+  CreateInferMeta(shape, dtype, layout, &meta_out);
   FullKernel<T, Context>(dev_ctx, shape, val, &dense_out);
   return dense_out;
 }
@@ -55,8 +56,9 @@ DenseTensor FullLike(
     DataType dtype = DataType::UNDEFINED,
     Backend backend = Backend::UNDEFINED,  // Is backend needed here?
     DataLayout layout = DataLayout::UNDEFINED) {
-  auto out_meta = CreateLikeInferMeta(x.meta(), dtype, layout);
-  auto dense_out = Empty<T, Context>(dev_ctx, std::move(out_meta));
+  auto dense_out = Empty<T, Context>(dev_ctx);
+  MetaTensor meta_out(&dense_out);
+  CreateLikeInferMeta(x, dtype, layout, &meta_out);
   FullLikeKernel<T, Context>(dev_ctx, val, &dense_out);
   return dense_out;
 }
diff --git a/paddle/pten/kernels/funcs/concat_funcs.h b/paddle/pten/kernels/funcs/concat_funcs.h
index 8455b8096922c3c6c78a20ac7cf05895c668f0ca..88fdad3a6da0d266f68714e6120921c1fcc54d0d 100644
--- a/paddle/pten/kernels/funcs/concat_funcs.h
+++ b/paddle/pten/kernels/funcs/concat_funcs.h
@@ -35,7 +35,7 @@ static inline int64_t ComputeAxis(int64_t axis, int64_t rank) {
 }
 
 static inline pten::DDim ComputeAndCheckShape(
-    const bool is_runtime,
+    bool is_runtime,
     const std::vector<pten::DDim>& inputs_dims,
     const size_t axis) {
   const size_t n = inputs_dims.size();
diff --git a/paddle/pten/kernels/math_kernel.h b/paddle/pten/kernels/math_kernel.h
index 6dcb9dd77adc2d978ba50a581b0a94f614ae1c28..eb39b618eb6b3a87fca30a9d27706979774bbbac 100644
--- a/paddle/pten/kernels/math_kernel.h
+++ b/paddle/pten/kernels/math_kernel.h
@@ -109,8 +109,9 @@ template <typename T, typename Context>
 DenseTensor Add(const Context& dev_ctx,
                 const DenseTensor& x,
                 const DenseTensor& y) {
-  auto out_meta = ElementwiseRawInferMeta(x.meta(), y.meta(), -1);
-  auto dense_out = pten::Empty<T, Context>(dev_ctx, std::move(out_meta));
+  auto dense_out = pten::Empty<T, Context>(dev_ctx);
+  MetaTensor meta_out(&dense_out);
+  ElementwiseInferMeta(x, y, &meta_out);
   AddKernel<T, Context>(dev_ctx, x, y, &dense_out);
   return dense_out;
 }
@@ -119,8 +120,9 @@ template <typename T, typename Context>
 DenseTensor Subtract(const Context& dev_ctx,
                      const DenseTensor& x,
                      const DenseTensor& y) {
-  auto out_meta = ElementwiseRawInferMeta(x.meta(), y.meta(), -1);
-  auto dense_out = pten::Empty<T, Context>(dev_ctx, std::move(out_meta));
+  auto dense_out = pten::Empty<T, Context>(dev_ctx);
+  MetaTensor meta_out(&dense_out);
+  ElementwiseInferMeta(x, y, &meta_out);
   SubtractKernel<T, Context>(dev_ctx, x, y, &dense_out);
   return dense_out;
 }
@@ -129,8 +131,9 @@ template <typename T, typename Context>
 DenseTensor Divide(const Context& dev_ctx,
                    const DenseTensor& x,
                    const DenseTensor& y) {
-  auto out_meta = ElementwiseRawInferMeta(x.meta(), y.meta(), -1);
-  auto dense_out = pten::Empty<T, Context>(dev_ctx, std::move(out_meta));
+  auto dense_out = pten::Empty<T, Context>(dev_ctx);
+  MetaTensor meta_out(&dense_out);
+  ElementwiseInferMeta(x, y, &meta_out);
   DivideKernel<T, Context>(dev_ctx, x, y, &dense_out);
   return dense_out;
 }
@@ -139,8 +142,9 @@ template <typename T, typename Context>
 DenseTensor Multiply(const Context& dev_ctx,
                      const DenseTensor& x,
                      const DenseTensor& y) {
-  auto out_meta = ElementwiseRawInferMeta(x.meta(), y.meta(), -1);
-  auto dense_out = pten::Empty<T, Context>(dev_ctx, std::move(out_meta));
+  auto dense_out = pten::Empty<T, Context>(dev_ctx);
+  MetaTensor meta_out(&dense_out);
+  ElementwiseInferMeta(x, y, &meta_out);
   MultiplyKernel<T, Context>(dev_ctx, x, y, &dense_out);
   return dense_out;
 }
@@ -150,8 +154,9 @@ DenseTensor Mean(const Context& dev_ctx,
                  const DenseTensor& x,
                  const std::vector<int64_t>& axis,
                  bool keep_dim) {
-  auto out_meta = ReduceInferMeta(x.meta(), axis, keep_dim);
-  auto dense_out = pten::Empty<T, Context>(dev_ctx, std::move(out_meta));
+  auto dense_out = pten::Empty<T, Context>(dev_ctx);
+  MetaTensor meta_out(&dense_out);
+  ReduceInferMeta(x, axis, keep_dim, x.dtype(), &meta_out);
   MeanKernel<T, Context>(dev_ctx, x, axis, keep_dim, &dense_out);
   return dense_out;
 }
@@ -162,9 +167,9 @@ DenseTensor Sum(const Context& dev_ctx,
                 const std::vector<int64_t>& axis,
                 DataType dtype,
                 bool keep_dim) {
-  auto out_meta = SumInferMeta(x.meta(), axis, dtype, keep_dim);
-  auto dense_out = pten::Empty<T, Context>(dev_ctx, std::move(out_meta));
-
+  auto dense_out = pten::Empty<T, Context>(dev_ctx);
+  MetaTensor meta_out(&dense_out);
+  SumInferMeta(x, axis, dtype, keep_dim, &meta_out);
   SumKernel<T, Context>(dev_ctx, x, axis, dtype, keep_dim, &dense_out);
   return dense_out;
 }
diff --git a/paddle/pten/kernels/matmul_kernel.h b/paddle/pten/kernels/matmul_kernel.h
index f9cb2c3801caa863b135ec2d4d188efff6dceb81..e6b9302cadd0ecfbcae3f31dc46de6184656efeb 100644
--- a/paddle/pten/kernels/matmul_kernel.h
+++ b/paddle/pten/kernels/matmul_kernel.h
@@ -35,8 +35,9 @@ DenseTensor Matmul(const Context& dev_ctx,
                    const DenseTensor& y,
                    bool transpose_x,
                    bool transpose_y) {
-  auto out_meta = MatmulInferMeta(x.meta(), y.meta(), transpose_x, transpose_y);
-  auto dense_out = Empty<T, Context>(dev_ctx, std::move(out_meta));
+  auto dense_out = Empty<T, Context>(dev_ctx);
+  MetaTensor meta_out(&dense_out);
+  MatmulInferMeta(x, y, transpose_x, transpose_y, &meta_out);
   MatmulKernel<T, Context>(dev_ctx, x, y, transpose_x, transpose_y, &dense_out);
   return dense_out;
 }
diff --git a/paddle/pten/kernels/reshape_kernel.cc b/paddle/pten/kernels/reshape_kernel.cc
index a76dfb09a0ea4ad4f531c323e015c8cdccdd165d..c52d251582bb508d5c47407d22c25137999c4182 100644
--- a/paddle/pten/kernels/reshape_kernel.cc
+++ b/paddle/pten/kernels/reshape_kernel.cc
@@ -26,15 +26,18 @@ void ReshapeKernel(const Context& dev_ctx,
                    const DenseTensor& x,
                    const ScalarArray& shape,
                    DenseTensor* out) {
-  auto out_meta = InferMetaFromVecValue(x.meta(), shape.GetData());
+  MetaTensor meta_out(out);
+  InferMetaFromVecValue(x, shape.GetData(), &meta_out);
   if (x.initialized() && x.Holder() == out->Holder()) {
-    out->ResizeAndAllocate(out_meta.dims);
+    dev_ctx.Alloc(out);
     return;
   }
-  out->set_meta(out_meta);
   dev_ctx.Alloc(out);
+  // TODO(chenweihang): the output dims are overwrite after copying,
+  // here we need to use copy method that only copy data
+  auto dims = out->dims();
   pten::Copy(dev_ctx, x, false, out);
-  out->Resize(out_meta.dims);
+  out->Resize(dims);
   out->ResetLoD(x.lod());
 }
 
diff --git a/paddle/pten/kernels/reshape_kernel.h b/paddle/pten/kernels/reshape_kernel.h
index 293f6cd2baf61dea1d9d40591690e2ac2ef76dac..a5672ad6e5b04a4c423f33ac5efef41d62b2423c 100644
--- a/paddle/pten/kernels/reshape_kernel.h
+++ b/paddle/pten/kernels/reshape_kernel.h
@@ -38,8 +38,9 @@ template <typename T, typename Context>
 DenseTensor Reshape(const Context& dev_ctx,
                     const DenseTensor& x,
                     const std::vector<int64_t>& shape) {
-  auto out_meta = InferMetaFromVecValue(x.meta(), shape);
-  auto dense_out = Empty<T, Context>(dev_ctx, std::move(out_meta));
+  auto dense_out = Empty<T, Context>(dev_ctx);
+  MetaTensor meta_out(&dense_out);
+  InferMetaFromVecValue(x, shape, &meta_out);
   ReshapeKernel<Context>(dev_ctx, x, ScalarArray(shape), &dense_out);
   return dense_out;
 }
diff --git a/paddle/pten/kernels/scale_kernel.h b/paddle/pten/kernels/scale_kernel.h
index 7fe627b4427444252db9ca1cd41ed02e1eae96e0..357bc70b40d98efffd31a352d230f69828565c4d 100644
--- a/paddle/pten/kernels/scale_kernel.h
+++ b/paddle/pten/kernels/scale_kernel.h
@@ -43,8 +43,9 @@ DenseTensor Scale(const Context& dev_ctx,
                   const Scalar& scale,
                   float bias,
                   bool bias_after_scale) {
-  auto out_meta = UnchangedInferMeta(x.meta());
-  auto dense_out = pten::Empty<T, Context>(dev_ctx, std::move(out_meta));
+  auto dense_out = pten::Empty<T, Context>(dev_ctx);
+  MetaTensor meta_out(&dense_out);
+  UnchangedInferMeta(x, &meta_out);
   ScaleKernel<T, Context>(
       dev_ctx, x, scale, bias, bias_after_scale, &dense_out);
   return dense_out;
diff --git a/paddle/pten/kernels/sign_kernel.h b/paddle/pten/kernels/sign_kernel.h
index 304b640d2af69f1f16dba6f6455656836d500516..4161c76e4719521daf3a4cec48109d778ad17301 100644
--- a/paddle/pten/kernels/sign_kernel.h
+++ b/paddle/pten/kernels/sign_kernel.h
@@ -25,8 +25,9 @@ void SignKernel(const Context& dev_ctx, const DenseTensor& x, DenseTensor* out);
 
 template <typename T, typename Context>
 DenseTensor Sign(const Context& dev_ctx, const DenseTensor& x) {
-  auto out_meta = UnchangedInferMeta(x.meta());
-  auto dense_out = pten::Empty<T, Context>(dev_ctx, std::move(out_meta));
+  auto dense_out = pten::Empty<T, Context>(dev_ctx);
+  MetaTensor meta_out(&dense_out);
+  UnchangedInferMeta(x, &meta_out);
   SignKernel<T, Context>(dev_ctx, x, &dense_out);
   return dense_out;
 }
diff --git a/paddle/pten/tests/api/scale_api.h b/paddle/pten/tests/api/scale_api.h
index 4b5a0a7daf4825c434670bf8b795f3f8598fed0e..b3b8b8f77c15abad503e4d608a904669b8458ab1 100644
--- a/paddle/pten/tests/api/scale_api.h
+++ b/paddle/pten/tests/api/scale_api.h
@@ -23,6 +23,7 @@
 #include "paddle/pten/common/scalar.h"
 #include "paddle/pten/common/scalar_array.h"
 #include "paddle/pten/core/kernel_registry.h"
+#include "paddle/pten/core/meta_tensor.h"
 #include "paddle/pten/infermeta/unary.h"
 #include "paddle/pten/kernels/scale_kernel.h"
 
@@ -68,11 +69,12 @@ PADDLE_API Tensor scale_kernel_context(const Tensor& x,
   kernel_context.EmplaceBackAttr(bias);
   kernel_context.EmplaceBackAttr(bias_after_scale);
 
-  auto out_meta = pten::UnchangedInferMeta(dense_x->meta());
   auto dense_out = std::make_shared<pten::DenseTensor>(
       pten::make_intrusive<paddle::experimental::SharedStorage>(
           pten::TransToFluidPlace(kernel_backend)),
-      std::move(out_meta));
+      pten::DenseTensorMeta());
+  pten::MetaTensor meta_out(dense_out.get());
+  pten::UnchangedInferMeta(*dense_x, &meta_out);
   kernel_context.EmplaceBackOutput(dense_out.get());
 
   Tensor out;
@@ -234,11 +236,12 @@ Tensor scale_switch_case(const Tensor& x,
 
   auto dense_x = std::dynamic_pointer_cast<pten::DenseTensor>(x.impl());
 
-  auto out_meta = pten::UnchangedInferMeta(dense_x->meta());
   auto dense_out = std::make_shared<pten::DenseTensor>(
       pten::make_intrusive<paddle::experimental::SharedStorage>(
           pten::TransToFluidPlace(kernel_backend)),
-      std::move(out_meta));
+      pten::DenseTensorMeta());
+  pten::MetaTensor meta_out(dense_out.get());
+  pten::UnchangedInferMeta(*dense_x, &meta_out);
 
   Tensor out;
   out.set_impl(dense_out);
diff --git a/python/paddle/utils/code_gen/api.yaml b/python/paddle/utils/code_gen/api.yaml
index e5ccd6b04054a27c6df9963735dc63c7db9c7c56..7768cb926e454cf8c5b10c097e1173f290fddff6 100644
--- a/python/paddle/utils/code_gen/api.yaml
+++ b/python/paddle/utils/code_gen/api.yaml
@@ -23,7 +23,7 @@
   output : Tensor
   infer_meta :
     func : ConcatInferMeta
-    param : [x, axis, true]
+    param : [x, axis]
   kernel :
     func : concat
 
diff --git a/python/paddle/utils/code_gen/api_gen.py b/python/paddle/utils/code_gen/api_gen.py
index a610bc9e806c8b5df87d455d3069587008955f04..cc7b31559f67c0ae21273a81688379b186a55ad5 100644
--- a/python/paddle/utils/code_gen/api_gen.py
+++ b/python/paddle/utils/code_gen/api_gen.py
@@ -65,13 +65,15 @@ PADDLE_API {self.return_type} {self.api}({self.args['args_declare']});
 
     def gene_output(self, output_type_list):
         kernel_output = ""
+        output_names = []
         output_create = ""
 
         if len(output_type_list) == 1:
             kernel_output = 'dense_out'
+            output_names.append('dense_out')
             output_create = f"""
   {self.return_type} out;
-  auto dense_out = SetKernelOutput(out_meta, kernel_backend, &out);"""
+  auto dense_out = SetKernelOutput(kernel_backend, &out);"""
 
         elif len(output_type_list) > 1:
             output_create = f"""
@@ -79,8 +81,9 @@ PADDLE_API {self.return_type} {self.api}({self.args['args_declare']});
 
             for i in range(len(output_type_list)):
                 kernel_output = kernel_output + f'dense_out_{i}, '
+                output_names.append(f'dense_out_{i}')
                 output_create = output_create + f"""
-  auto dense_out_{i} = SetKernelOutput(std::get<{i}>(out_meta), kernel_backend, &std::get<{i}>(out));"""
+  auto dense_out_{i} = SetKernelOutput(kernel_backend, &std::get<{i}>(out));"""
 
             kernel_output = kernel_output[:-2]
         else:
@@ -88,22 +91,23 @@ PADDLE_API {self.return_type} {self.api}({self.args['args_declare']});
                 "{} : Output error: the output should not be empty.".format(
                     self.api))
 
-        return kernel_output, output_create
+        return kernel_output, output_names, output_create
 
     def gene_api_code(self):
         if self.is_base_api:
             input_tensors, kernel_args, kernel_signature = gen_utils.get_kernel_args(
                 self.args['inputs'], self.args['attrs'], self.out_type_list,
                 self.kernel['param'])
-            outputs_args, output_create = self.gene_output(self.out_type_list)
+            outputs_args, output_names, output_create = self.gene_output(
+                self.out_type_list)
             return f"""
 PADDLE_API {self.return_type} {self.api}({self.args["args_define"]}) {{
 {gen_utils.gene_kernel_select(self.api, self.args['inputs']['names'], self.args['attrs'], self.kernel)}
 
   auto* dev_ctx = GetDeviceContextByBackend(kernel_backend);
 {input_tensors}
-{gen_utils.gene_infer_meta(self.args['inputs']['names'], self.args['attrs']['names'], self.infer_meta)}
 {output_create}
+{gen_utils.gene_infer_meta(self.args['inputs']['names'], self.args['attrs']['names'], output_names, self.infer_meta)}
   using kernel_signature = {kernel_signature};
   auto* kernel_fn = kernel.GetVariadicKernelFn<kernel_signature>();
   (*kernel_fn)({kernel_args}, {outputs_args});
diff --git a/python/paddle/utils/code_gen/backward_api_gen.py b/python/paddle/utils/code_gen/backward_api_gen.py
index e53886339c48d28641e90b3b6c9a21a2b8615cc4..9df25e2dead0b5222a92028bbac683dbedc0e158 100644
--- a/python/paddle/utils/code_gen/backward_api_gen.py
+++ b/python/paddle/utils/code_gen/backward_api_gen.py
@@ -105,13 +105,15 @@ class BackwardAPI:
 
     def gene_output(self, output_type_list):
         kernel_output = ""
+        output_names = []
         output_create = ""
 
         if len(output_type_list) == 1:
             kernel_output = 'dense_out'
+            output_names.append('dense_out')
             output_create = f"""
   {self.return_type} out;
-  auto dense_out = SetKernelOutput(out_meta, kernel_backend, &out);"""
+  auto dense_out = SetKernelOutput(kernel_backend, &out);"""
 
         elif len(output_type_list) > 1:
             output_create = f"""
@@ -119,6 +121,7 @@ class BackwardAPI:
 
             for i, out_type_item in enumerate(output_type_list):
                 kernel_output = kernel_output + f'dense_out_{i}, '
+                output_names.append(f'dense_out_{i}')
                 if out_type_item == 'Tensor':
                     get_out_code = f'&out[{i}][0]'
                     output_create = output_create + f"""
@@ -127,7 +130,7 @@ class BackwardAPI:
                 else:
                     get_out_code = f'&out[{i}]'
                 output_create = output_create + f"""
-  auto dense_out_{i} = SetKernelOutput(std::get<{i}>(out_meta), kernel_backend, {get_out_code});"""
+  auto dense_out_{i} = SetKernelOutput(kernel_backend, {get_out_code});"""
 
             kernel_output = kernel_output[:-2]
         else:
@@ -135,14 +138,14 @@ class BackwardAPI:
                 "{} : Output error: the output should not be empty.".format(
                     self.backward_api))
 
-        return kernel_output, output_create
+        return kernel_output, output_names, output_create
 
     def gene_api_code(self):
         if self.is_base_api:
             input_tensors, kernel_args, kernel_signature = gen_utils.get_kernel_args(
                 self.args['inputs'], self.args['attrs'], self.output_type_list,
                 self.kernel['param'])
-            outputs_args, output_create = self.gene_output(
+            outputs_args, output_names, output_create = self.gene_output(
                 self.output_type_list)
             return f"""
 // {self.return_comment}
@@ -151,8 +154,8 @@ class BackwardAPI:
 
   auto* dev_ctx = GetDeviceContextByBackend(kernel_backend);
 {input_tensors}
-{gen_utils.gene_infer_meta(self.args['inputs']['names'], self.args['attrs']['names'], self.infer_meta)}
 {output_create}
+{gen_utils.gene_infer_meta(self.args['inputs']['names'], self.args['attrs']['names'], output_names, self.infer_meta)}
 
   using kernel_signature = {kernel_signature};
   auto* kernel_fn = kernel.GetVariadicKernelFn<kernel_signature>();
diff --git a/python/paddle/utils/code_gen/gen_utils.py b/python/paddle/utils/code_gen/gen_utils.py
index f1aa299fb17cf4ba1583974dcdf594942b5f97e5..56143a8f517cfd64468377070f5b91eff0502fef 100644
--- a/python/paddle/utils/code_gen/gen_utils.py
+++ b/python/paddle/utils/code_gen/gen_utils.py
@@ -15,6 +15,7 @@
 import re
 
 PREFIX_TENSOR_NAME = 'dense_'
+PREFIX_META_TENSOR_NAME = 'meta_'
 
 
 def parse_args(api_name, args_str):
@@ -265,13 +266,21 @@ def gene_kernel_select(api, input_names, attrs, kernel) -> str:
     return kernel_select_code
 
 
-def gene_infer_meta(input_names, attr_names, infer_meta) -> str:
-    infer_meta_params = infer_meta['param'] if infer_meta[
-        'param'] is not None else input_names + attr_names
+def gene_infer_meta(input_names, attr_names, output_names, infer_meta) -> str:
+    infer_meta_params = infer_meta['param'] + output_names if infer_meta[
+        'param'] is not None else input_names + attr_names + output_names
+    # generate meta tensors
+    meta_tensor_code = ""
     param_code = ""
     for param in infer_meta_params:
         if param in input_names:
-            param_code = param_code + "GetDenseTensorMeta(*" + PREFIX_TENSOR_NAME + param + "), "
+            param_code = param_code + "MakeMetaTensor(*" + PREFIX_TENSOR_NAME + param + "), "
+        elif param in output_names:
+            meta_tensor_code = meta_tensor_code + "  pten::MetaTensor " + param.replace(
+                PREFIX_TENSOR_NAME,
+                PREFIX_META_TENSOR_NAME) + "(" + param + ");\n"
+            param_code = param_code + "&" + param.replace(
+                PREFIX_TENSOR_NAME, PREFIX_META_TENSOR_NAME) + ", "
         elif param in attr_names:
             param_code = param_code + param + ", "
         elif isinstance(param, str):
@@ -282,8 +291,8 @@ def gene_infer_meta(input_names, attr_names, infer_meta) -> str:
             param_code = param_code + str(param) + ", "
 
     param_code = param_code[:-2]
-    return f"""
-  auto out_meta = pten::{infer_meta['func']}({param_code});
+    return f"""{meta_tensor_code}
+  pten::{infer_meta['func']}({param_code});
 """