diff --git a/paddle/fluid/imperative/tracer.cc b/paddle/fluid/imperative/tracer.cc
index b20e27bebbeddf8cf4bfa868716458d8c0b11a87..f1374bc8f7bd7cacb1c96575a6b3472b1cc3d9b4 100644
--- a/paddle/fluid/imperative/tracer.cc
+++ b/paddle/fluid/imperative/tracer.cc
@@ -410,6 +410,9 @@ void Tracer::TraceOp(const std::string& type,
   VLOG(6) << "Running On Eager TraceOp with use_default_attr_map: "
           << use_default_attr_map;
   std::map<phi::DenseTensor*, phi::DenseTensor*> need_backup_inputs2outputs;
+  std::map<phi::DenseTensor*, std::shared_ptr<phi::Allocation>>
+      need_backup_inputs2holder;
+  std::map<phi::DenseTensor*, phi::DDim> need_backup_inputs2strides;
   if (FLAGS_use_stride_kernel) {
     for (auto& iter : inplace_map) {
       auto inputs_iter = ins.find(iter.first);
@@ -426,11 +429,12 @@ void Tracer::TraceOp(const std::string& type,
                 outputs_iter->second[i]
                     ->MutableVar()
                     ->GetMutable<phi::DenseTensor>();
+            need_backup_inputs2holder[dense_tensor] = dense_tensor->Holder();
+            need_backup_inputs2strides[dense_tensor] = dense_tensor->strides();
           }
         }
       }
     }
-
     TraceOpImpl<egr::EagerVariable>(type,
                                     ins,
                                     outs,
@@ -443,7 +447,11 @@ void Tracer::TraceOp(const std::string& type,
 
     auto dev_ctx = paddle::platform::DeviceContextPool::Instance().Get(place);
     for (auto& iter : need_backup_inputs2outputs) {
-      paddle::experimental::TransStride(dev_ctx, iter.second, iter.first);
+      iter.first->ResetHolder(need_backup_inputs2holder[iter.first]);
+      iter.first->set_strides(need_backup_inputs2strides[iter.first]);
+      paddle::experimental::TransStrideLegacy(dev_ctx, iter.second, iter.first);
+      iter.second->ResetHolder(need_backup_inputs2holder[iter.first]);
+      iter.second->set_strides(need_backup_inputs2strides[iter.first]);
     }
   } else {
     TraceOpImpl<egr::EagerVariable>(type,
diff --git a/paddle/phi/api/lib/api_gen_utils.cc b/paddle/phi/api/lib/api_gen_utils.cc
index d1e549c91c4df17c432ef02cc8239bf6bc96a7df..e3f58683d988a7898437c2b0ec478860a830c2d2 100644
--- a/paddle/phi/api/lib/api_gen_utils.cc
+++ b/paddle/phi/api/lib/api_gen_utils.cc
@@ -423,6 +423,56 @@ void TransStride(phi::DeviceContext* dev_ctx,
   }
 }
 
+void TransStrideLegacy(phi::DeviceContext* dev_ctx,
+                       phi::DenseTensor* from,
+                       phi::DenseTensor* to) {
+  if (to) {
+    auto* cpu_ctx = dynamic_cast<phi::CPUContext*>(dev_ctx);
+    if (cpu_ctx) {
+      PD_VISIT_ALL_TYPES(to->dtype(), "StridedCopyKernel", ([&] {
+                           phi::StridedCopyKernel<data_t, phi::CPUContext>(
+                               *cpu_ctx,
+                               *from,
+                               phi::vectorize<int64_t>(to->dims()),
+                               phi::vectorize<int64_t>(to->strides()),
+                               to->offset(),
+                               to);
+                         }));
+      return;
+    }
+#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
+    auto* gpu_ctx = dynamic_cast<phi::GPUContext*>(dev_ctx);
+    if (gpu_ctx) {
+      PD_VISIT_ALL_TYPES(to->dtype(), "StridedCopyKernel", ([&] {
+                           phi::StridedCopyKernel<data_t, phi::GPUContext>(
+                               *gpu_ctx,
+                               *from,
+                               phi::vectorize<int64_t>(to->dims()),
+                               phi::vectorize<int64_t>(to->strides()),
+                               to->offset(),
+                               to);
+                         }));
+      return;
+    }
+#endif
+#ifdef PADDLE_WITH_XPU
+    auto* xpu_ctx = dynamic_cast<phi::XPUContext*>(dev_ctx);
+    if (xpu_ctx) {
+      PD_VISIT_ALL_TYPES(to->dtype(), "StridedCopyKernel", ([&] {
+                           phi::StridedCopyKernel<data_t, phi::XPUContext>(
+                               *xpu_ctx,
+                               *from,
+                               phi::vectorize<int64_t>(to->dims()),
+                               phi::vectorize<int64_t>(to->strides()),
+                               to->offset(),
+                               to);
+                         }));
+      return;
+    }
+#endif
+  }
+}
+
 void TransStride(phi::DeviceContext* dev_ctx,
                  const std::vector<phi::DenseTensor*>& from,
                  const std::vector<phi::DenseTensor*>& to) {
diff --git a/paddle/phi/api/lib/api_gen_utils.h b/paddle/phi/api/lib/api_gen_utils.h
index afe312b7096416a262ede2256906d69479470a24..1b552bf94eaa4d51eddd3955752a9a7f6a5a48f7 100644
--- a/paddle/phi/api/lib/api_gen_utils.h
+++ b/paddle/phi/api/lib/api_gen_utils.h
@@ -133,6 +133,10 @@ void TransStride(phi::DeviceContext* dev_ctx,
                  phi::SelectedRows* from,
                  phi::SelectedRows* to);
 
+void TransStrideLegacy(phi::DeviceContext* dev_ctx,
+                       phi::DenseTensor* from,
+                       phi::DenseTensor* to);
+
 #ifdef PADDLE_WITH_DISTRIBUTE
 /* ------------------ for auto parallel ----------------------- */