Patch del (#52754)

* [DO NOT MERGE] adadelta lr support

* [DO NOT MERGE] gpu support

* [test] follow torch

* fix acc update order

* for ci

* [bug fix] update master para

* [bug fix] update test

* [bug fix] for ci test

* for ci

* fix xpu

* [adadelta fix] del fluid head file

* for ci

* del notes

paddle/phi/kernels/impl/adadelta_kernel_impl.h

@@ -13,10 +13,6 @@
 // limitations under the License.
 
 #pragma once
-#include "paddle/fluid/framework/eigen.h"
-#include "paddle/fluid/framework/op_registry.h"
-#include "paddle/fluid/operators/eigen/eigen_function.h"
-
 #include "paddle/phi/common/amp_type_traits.h"
 #include "paddle/phi/kernels/adadelta_kernel.h"
 #include "paddle/phi/kernels/funcs/eigen/common.h"
@@ -67,26 +63,20 @@ void AdadeltaKernel(const Context& dev_ctx,
       -(((eigen_avg_squared_update + epsilon_).sqrt()) /
         ((eigen_avg_squared_grad_out + epsilon_).sqrt()) * eigen_grad_cast);
   Eigen::DSizes<int, 1> m_dsize(avg_squared_update_out->numel());
-  if (paddle::platform::is_cpu_place(dev_ctx.GetPlace())) {
-    auto* lr = learning_rate.data<T>();
+  auto lr = EigenVector<MPDType>::Flatten(learning_rate);
+  if (multi_precision) {
+    auto eigen_master_param_out =
+        EigenVector<MPDType>::Flatten(*master_param_outs);
+    auto eigen_master_param = EigenVector<MPDType>::Flatten(*master_param);
+
+    eigen_master_param_out.device(place) =
+        eigen_master_param + lr.broadcast(m_dsize) * update;
     eigen_param_out.device(place) =
-        eigen_param + lr[0] * update.template cast<T>();
+        (eigen_param.template cast<MPDType>() + lr.broadcast(m_dsize) * update)
+            .template cast<T>();
   } else {
-    auto lr = EigenVector<MPDType>::Flatten(learning_rate);
-    if (multi_precision) {
-      auto eigen_master_param_out =
-          EigenVector<MPDType>::Flatten(*master_param_outs);
-      auto eigen_master_param = EigenVector<MPDType>::Flatten(*master_param);
-
-      eigen_master_param_out.device(place) =
-          eigen_master_param + lr.broadcast(m_dsize) * update;
-      eigen_param_out.device(place) = (eigen_param.template cast<MPDType>() +
-                                       lr.broadcast(m_dsize) * update)
-                                          .template cast<T>();
-    } else {
-      eigen_param_out.device(place) =
-          eigen_param + (lr.broadcast(m_dsize) * update).template cast<T>();
-    }
+    eigen_param_out.device(place) =
+        eigen_param + (lr.broadcast(m_dsize) * update).template cast<T>();
   }
   eigen_avg_squared_update_out.device(place) =
       rho_ * eigen_avg_squared_update + (1 - rho_) * update.square();