diff --git a/paddle/fluid/operators/group_norm_op.cu b/paddle/fluid/operators/group_norm_op.cu
index 12a989bc82b1ca1ed2a472db41385e3ba181310c..08ea4d3278b2312afb23117fed7c5dccb8eb3216 100644
--- a/paddle/fluid/operators/group_norm_op.cu
+++ b/paddle/fluid/operators/group_norm_op.cu
@@ -324,7 +324,7 @@ class GroupNormKernel<phi::GPUContext, T> : public framework::OpKernel<T> {
     dim3 grid(group_size, groups, x_dims[0]);
     dim3 threads(block_size, 1, 1);
     if (data_layout == DataLayout::kNCHW) {
-      using AccT = typename details::MPTypeTrait<T>::Type;
+      using AccT = typename phi::dtype::MPTypeTrait<T>::Type;
       constexpr int vec_size = sizeof(float4) / sizeof(T);
       int size = group_size * imsize;
       const int max_num_threads = 1024;
diff --git a/paddle/fluid/operators/uniform_random_op.h b/paddle/fluid/operators/uniform_random_op.h
index bf2666deda28b2be72c2bddd72c35a90f224a988..3ddf6092f04bff181387cbc605096c8ab3d68c51 100644
--- a/paddle/fluid/operators/uniform_random_op.h
+++ b/paddle/fluid/operators/uniform_random_op.h
@@ -165,7 +165,7 @@ void UniformRandom(const framework::ExecutionContext& context,
 
   if (seed == 0) {
     // Use global Generator seed
-    using MT = typename details::MPTypeTrait<T>::Type;
+    using MT = typename phi::dtype::MPTypeTrait<T>::Type;
     phi::funcs::uniform_distribution<MT> dist;
     phi::funcs::uniform_real_transform<MT> trans(min, max);
     phi::funcs::distribution_and_transform<T>(dev_cxt, tensor, dist, trans);
diff --git a/paddle/phi/kernels/funcs/functors.h b/paddle/phi/kernels/funcs/functors.h
index 2e6fe8b2d738bc6de9ea755c3fbd77af8c387513..3c7ae5ed09af3ca3ac177a166d81857f9df3f33d 100644
--- a/paddle/phi/kernels/funcs/functors.h
+++ b/paddle/phi/kernels/funcs/functors.h
@@ -14,7 +14,7 @@ limitations under the License. */
 
 #pragma once
 
-#include "paddle/fluid/operators/amp/fp16_type_traits.h"
+#include "paddle/phi/common/amp_type_traits.h"
 #include "paddle/phi/kernels/funcs/math.h"
 
 namespace phi {
@@ -38,7 +38,7 @@ struct AddGradFunctor {
 
 template <typename T>
 struct ScaleFunctor {
-  using MT = typename paddle::operators::details::MPTypeTrait<T>::Type;
+  using MT = typename phi::dtype::MPTypeTrait<T>::Type;
   explicit ScaleFunctor(const MT coeff) : coeff_(coeff) {}
 
   inline HOSTDEVICE T operator()(T ele) {
@@ -125,7 +125,7 @@ struct SigmoidGradFunctor {
 
 template <typename T>
 struct GeluFunctor {
-  using MT = typename paddle::operators::details::MPTypeTrait<T>::Type;
+  using MT = typename phi::dtype::MPTypeTrait<T>::Type;
   inline HOSTDEVICE T operator()(T x) {
     // this function is tanh approximation of gelu
     // actual gelu is:
@@ -141,7 +141,7 @@ struct GeluFunctor {
 
 template <typename T>
 struct GeluGradFunctor {
-  using MT = typename paddle::operators::details::MPTypeTrait<T>::Type;
+  using MT = typename phi::dtype::MPTypeTrait<T>::Type;
   inline HOSTDEVICE T UseX(T x) {
     MT mx = static_cast<MT>(x);
     MT tanh_out =
diff --git a/paddle/phi/kernels/gpu/norm_grad_kernel.cu b/paddle/phi/kernels/gpu/norm_grad_kernel.cu
index bbbb6e9c018b61b590ba934b78ecffc7ee111a6e..cb02cc713852c8e5c96edf0e82b02c183ad51b6c 100644
--- a/paddle/phi/kernels/gpu/norm_grad_kernel.cu
+++ b/paddle/phi/kernels/gpu/norm_grad_kernel.cu
@@ -22,8 +22,8 @@
 #include <hipcub/hipcub.hpp>
 namespace cub = hipcub;
 #endif
-#include "paddle/fluid/operators/amp/fp16_type_traits.h"
 #include "paddle/phi/backends/gpu/gpu_context.h"
+#include "paddle/phi/common/amp_type_traits.h"
 #include "paddle/phi/common/bfloat16.h"
 #include "paddle/phi/core/kernel_registry.h"
 #include "paddle/phi/kernels/funcs/common_shape.h"
@@ -38,7 +38,7 @@ __global__ void NormalizeGradient(const T* x,
                                   const int axis_n,
                                   const int post,
                                   T* x_grad) {
-  using MT = typename paddle::operators::details::MPTypeTrait<T>::Type;
+  using MT = typename phi::dtype::MPTypeTrait<T>::Type;
   typedef cub::BlockReduce<MT, BlockDim> BlockReduce;
   __shared__ typename BlockReduce::TempStorage temp_storage_sum;
   int num = pre * post;
diff --git a/paddle/phi/kernels/gpu/norm_kernel.cu b/paddle/phi/kernels/gpu/norm_kernel.cu
index bd9cffe79614b281dd3fa3d14501eec3922500c2..4843831ebfc683c715b0328a03b14614a86b2b2d 100644
--- a/paddle/phi/kernels/gpu/norm_kernel.cu
+++ b/paddle/phi/kernels/gpu/norm_kernel.cu
@@ -22,8 +22,8 @@
 #include <hipcub/hipcub.hpp>
 namespace cub = hipcub;
 #endif
-#include "paddle/fluid/operators/amp/fp16_type_traits.h"
 #include "paddle/phi/backends/gpu/gpu_context.h"
+#include "paddle/phi/common/amp_type_traits.h"
 #include "paddle/phi/common/float16.h"
 #include "paddle/phi/core/kernel_registry.h"
 #include "paddle/phi/kernels/funcs/common_shape.h"
@@ -46,7 +46,7 @@ __global__ void Normalize(const T* x,
                           const T eps,
                           T* y,
                           T* out_norm) {
-  using MT = typename paddle::operators::details::MPTypeTrait<T>::Type;
+  using MT = typename phi::dtype::MPTypeTrait<T>::Type;
   typedef cub::BlockReduce<MT, BlockDim> BlockReduce;
   __shared__ typename BlockReduce::TempStorage temp_storage;
   int num = pre * post;
diff --git a/paddle/phi/kernels/gpu/sgd_kernel.cu b/paddle/phi/kernels/gpu/sgd_kernel.cu
index ea257ebd1cc24fe59892574de5e27e37daa51e84..e3f0bf968c82cc754b31b07732f8e2314d6f2830 100644
--- a/paddle/phi/kernels/gpu/sgd_kernel.cu
+++ b/paddle/phi/kernels/gpu/sgd_kernel.cu
@@ -15,10 +15,10 @@
 #include "paddle/phi/kernels/sgd_kernel.h"
 
 #include "paddle/fluid/framework/mixed_vector.h"
-#include "paddle/fluid/operators/amp/fp16_type_traits.h"
 #include "paddle/phi/backends/gpu/gpu_context.h"
 #include "paddle/phi/backends/gpu/gpu_helper.h"
 #include "paddle/phi/backends/gpu/gpu_primitives.h"
+#include "paddle/phi/common/amp_type_traits.h"
 #include "paddle/phi/core/kernel_registry.h"
 
 namespace phi {
@@ -72,7 +72,7 @@ void SGDDenseKernel(const Context& dev_ctx,
                     bool multi_precision,
                     DenseTensor* param_out,
                     DenseTensor* master_param_out) {
-  using MPDType = typename paddle::operators::details::MPTypeTrait<T>::Type;
+  using MPDType = typename phi::dtype::MPTypeTrait<T>::Type;
   // do check here
   // if (multi_precision) {
   //   bool has_master =
@@ -109,7 +109,7 @@ void SGDDenseParamSparseGradKernel(
     bool multi_precision,
     DenseTensor* param_out,
     DenseTensor* master_param_out) {
-  using MPDType = typename paddle::operators::details::MPTypeTrait<T>::Type;
+  using MPDType = typename phi::dtype::MPTypeTrait<T>::Type;
   // do some check here
   // if (multi_precision) {
   //   bool has_master =
diff --git a/paddle/phi/kernels/primitive/functor_primitives.h b/paddle/phi/kernels/primitive/functor_primitives.h
index 700ba00088517236729ca6ee982111f507ad0c6c..b0f3d62823c09ca63f1bdfe9806ad68700016258 100644
--- a/paddle/phi/kernels/primitive/functor_primitives.h
+++ b/paddle/phi/kernels/primitive/functor_primitives.h
@@ -14,7 +14,7 @@
 
 #pragma once
 
-#include "paddle/fluid/operators/amp/fp16_type_traits.h"
+#include "paddle/phi/common/amp_type_traits.h"
 #include "paddle/phi/common/float16.h"
 #include "paddle/phi/core/enforce.h"
 #include "paddle/phi/kernels/funcs/eigen/extensions.h"
@@ -79,7 +79,7 @@ struct IdentityFunctor {
 template <typename Tx, typename Ty = Tx>
 struct DivideFunctor {
  private:
-  using MPType = typename ::paddle::operators::details::MPTypeTrait<Tx>::Type;
+  using MPType = typename ::phi::dtype::MPTypeTrait<Tx>::Type;
 
  public:
   HOSTDEVICE inline DivideFunctor() { n_inv = static_cast<MPType>(1.0f); }