diff --git a/paddle/pten/kernels/empty_kernel.h b/paddle/pten/kernels/empty_kernel.h
index 8a7da8fbd5659fdfb5af73e03e1cce8868b9c726..98f7e03d771bee306030a554633f9a2355747154 100644
--- a/paddle/pten/kernels/empty_kernel.h
+++ b/paddle/pten/kernels/empty_kernel.h
@@ -32,7 +32,7 @@ void EmptyLikeKernel(const Context& dev_ctx, DenseTensor* out);
 
 // TODO(chenweihang): the tensor creation method need to be replaced later,
 // all kernel api call Empty here instead of making tensor self
-template <typename T, typename Context>
+template <typename Context>
 DenseTensor Empty(const Context& dev_ctx, DenseTensorMeta&& meta) {
   pten::DenseTensor dense_out(
       pten::make_intrusive<paddle::experimental::SharedStorage>(
@@ -43,10 +43,10 @@ DenseTensor Empty(const Context& dev_ctx, DenseTensorMeta&& meta) {
 
 template <typename T, typename Context>
 DenseTensor Empty(const Context& dev_ctx) {
-  return Empty<T, Context>(dev_ctx,
-                           {paddle::experimental::CppTypeToDataType<T>::Type(),
-                            {-1},
-                            DataLayout::NCHW});
+  return Empty(dev_ctx,
+               {paddle::experimental::CppTypeToDataType<T>::Type(),
+                {-1},
+                DataLayout::NCHW});
 }
 
 template <typename T, typename Context>
diff --git a/paddle/pten/kernels/funcs/elementwise_base.h b/paddle/pten/kernels/funcs/elementwise_base.h
index 21abf82c7b2283da1a82f18c22d31335ee4d3335..0f26f3d8aa6619b8804ca2782a9b6147f32165f9 100644
--- a/paddle/pten/kernels/funcs/elementwise_base.h
+++ b/paddle/pten/kernels/funcs/elementwise_base.h
@@ -392,7 +392,7 @@ static inline void GetDoubleGradSafeTensor(const DeviceContext &dev_ctx,
     *ddx_safe = *ddx;
   } else {
     auto meta = pten::DenseTensorMeta(x.dtype(), x.dims(), x.layout());
-    *ddx_safe = pten::Empty<T, DeviceContext>(dev_ctx, std::move(meta));
+    *ddx_safe = pten::Empty(dev_ctx, std::move(meta));
     ddx_safe->mutable_data(dev_ctx.GetPlace());
     paddle::operators::math::SetConstant<DeviceContext, T> set_zero;
     set_zero(dev_ctx, ddx_safe, static_cast<T>(0));
diff --git a/paddle/pten/kernels/sparse/cpu/sparse_utils_kernel.cc b/paddle/pten/kernels/sparse/cpu/sparse_utils_kernel.cc
index de648f76c1bdef7f83632a42b8d4efcbd3bda24f..a86804e25dcdd644e4b4df1996914b34d20ebfbe 100644
--- a/paddle/pten/kernels/sparse/cpu/sparse_utils_kernel.cc
+++ b/paddle/pten/kernels/sparse/cpu/sparse_utils_kernel.cc
@@ -76,10 +76,8 @@ void DenseToSparseCooKernel(const Context& dev_ctx,
                                {sparse_dim, static_cast<int64_t>(non_zero_num)},
                                DataLayout::NCHW);
   DenseTensorMeta values_meta(x.meta().dtype, values_dims, x.meta().layout);
-  pten::DenseTensor indices =
-      pten::Empty<int64_t, Context>(dev_ctx, std::move(indices_meta));
-  pten::DenseTensor values =
-      pten::Empty<T, Context>(dev_ctx, std::move(values_meta));
+  pten::DenseTensor indices = pten::Empty(dev_ctx, std::move(indices_meta));
+  pten::DenseTensor values = pten::Empty(dev_ctx, std::move(values_meta));
   int64_t* indices_data = indices.mutable_data<int64_t>(place);
   T* values_data = values.mutable_data<T>(place);
 
@@ -123,10 +121,8 @@ void SparseCsrToCooKernel(const Context& dev_ctx,
   DenseTensorMeta indices_meta(
       DataType::INT64, {sparse_dim, non_zero_num}, DataLayout::NCHW);
   DenseTensorMeta values_meta(x.dtype(), {non_zero_num}, x.layout());
-  pten::DenseTensor indices =
-      pten::Empty<int64_t, Context>(dev_ctx, std::move(indices_meta));
-  pten::DenseTensor values =
-      pten::Empty<T, Context>(dev_ctx, std::move(values_meta));
+  pten::DenseTensor indices = pten::Empty(dev_ctx, std::move(indices_meta));
+  pten::DenseTensor values = pten::Empty(dev_ctx, std::move(values_meta));
   int64_t* coo_indices = indices.mutable_data<int64_t>(place);
   int64_t* batch_ptr = x_dims.size() == 2 ? nullptr : coo_indices;
   int64_t* coo_rows_data =
diff --git a/paddle/pten/kernels/sparse/gpu/sparse_utils_kernel.cu b/paddle/pten/kernels/sparse/gpu/sparse_utils_kernel.cu
index cb8e30753107e2dbfc08ad1fa82b4dd4aea417e7..cd3124bd02fcbedc0052c37c2cefe10b4d73a747 100644
--- a/paddle/pten/kernels/sparse/gpu/sparse_utils_kernel.cu
+++ b/paddle/pten/kernels/sparse/gpu/sparse_utils_kernel.cu
@@ -111,14 +111,12 @@ void DenseToSparseCooKernel(const Context& dev_ctx,
   const int cols = dims_2d[1];
   auto nums_meta =
       pten::DenseTensorMeta(DataType::INT32, {1}, pten::DataLayout::NCHW);
-  DenseTensor nums =
-      pten::Empty<int64_t, Context>(dev_ctx, std::move(nums_meta));
+  DenseTensor nums = pten::Empty(dev_ctx, std::move(nums_meta));
   auto x_dims_meta =
       pten::DenseTensorMeta(DataType::INT64,
                             {static_cast<int64_t>(x_dims.size())},
                             pten::DataLayout::NCHW);
-  DenseTensor d_x_dims =
-      pten::Empty<T, Context>(dev_ctx, std::move(x_dims_meta));
+  DenseTensor d_x_dims = pten::Empty(dev_ctx, std::move(x_dims_meta));
 
   const auto place = dev_ctx.GetPlace();
 
@@ -136,8 +134,7 @@ void DenseToSparseCooKernel(const Context& dev_ctx,
 
   auto temp_indexs_meta =
       pten::DenseTensorMeta(DataType::INT32, {rows}, pten::DataLayout::NCHW);
-  DenseTensor temp_indexs =
-      pten::Empty<T, Context>(dev_ctx, std::move(temp_indexs_meta));
+  DenseTensor temp_indexs = pten::Empty(dev_ctx, std::move(temp_indexs_meta));
   int* temp_indexs_ptr = temp_indexs.mutable_data<int>(place);
   GetNonZeroNums<<<grid_size, block_size, 0, dev_ctx.stream()>>>(
       x_data, rows, cols, nums_ptr, temp_indexs_ptr);
@@ -269,11 +266,9 @@ void SparseCsrToCooKernel(const Context& dev_ctx,
       DataType::INT64, {sparse_dim, non_zero_num}, DataLayout::NCHW);
   DenseTensorMeta values_meta(x.dtype(), {non_zero_num}, x.layout());
   DenseTensorMeta offsets_meta(DataType::INT32, {batchs}, DataLayout::NCHW);
-  DenseTensor indices =
-      pten::Empty<int64_t, Context>(dev_ctx, std::move(indices_meta));
-  DenseTensor values = pten::Empty<T, Context>(dev_ctx, std::move(values_meta));
-  DenseTensor offsets =
-      pten::Empty<T, Context>(dev_ctx, std::move(offsets_meta));
+  DenseTensor indices = pten::Empty(dev_ctx, std::move(indices_meta));
+  DenseTensor values = pten::Empty(dev_ctx, std::move(values_meta));
+  DenseTensor offsets = pten::Empty(dev_ctx, std::move(offsets_meta));
   int64_t* coo_indices = indices.mutable_data<int64_t>(place);
   int64_t* batch_ptr = x_dims.size() == 2 ? nullptr : coo_indices;
   int64_t* coo_rows_data =