PaddlePaddle
diff --git a/‎paddle/phi/api/lib/api_gen_utils.cc
Copy file name to clipboard
+23Lines changed: 23 additions & 0 deletions b/‎paddle/phi/api/lib/api_gen_utils.cc
Copy file name to clipboard
+23Lines changed: 23 additions & 0 deletions
diff --git a/‎paddle/phi/api/lib/api_gen_utils.h
Copy file name to clipboard
+6Lines changed: 6 additions & 0 deletions b/‎paddle/phi/api/lib/api_gen_utils.h
Copy file name to clipboard
+6Lines changed: 6 additions & 0 deletions
diff --git a/‎paddle/phi/api/lib/data_transform.cc
Copy file name to clipboard
+34-10Lines changed: 34 additions & 10 deletions b/‎paddle/phi/api/lib/data_transform.cc
Copy file name to clipboard
+34-10Lines changed: 34 additions & 10 deletions
diff --git a/‎paddle/phi/api/lib/data_transform.h
Copy file name to clipboard
+12Lines changed: 12 additions & 0 deletions b/‎paddle/phi/api/lib/data_transform.h
Copy file name to clipboard
+12Lines changed: 12 additions & 0 deletions
@@ -590,5 +590,28 @@ std::vector<phi::distributed::DistTensor*> SetKernelDistOutput(
   return results;
 }
 
+std::vector<phi::distributed::DistTensor*> SetKernelDistInplaceOutput(
+    size_t out_size, std::vector<Tensor>* out) {
+  std::vector<phi::distributed::DistTensor*> results(out->size(), nullptr);
+  for (size_t i = 0; i < out->size(); ++i) {
+    results[i] =
+        static_cast<phi::distributed::DistTensor*>(out->at(i).impl().get());
+  }
+  return results;
+}
+
+std::vector<phi::distributed::DistTensor*> SetKernelDistInplaceOptionalOutput(
+    size_t out_size, paddle::optional<std::vector<Tensor>> out) {
+  std::vector<phi::distributed::DistTensor*> results;
+  if (out) {
+    results = std::vector<phi::distributed::DistTensor*>(out->size(), nullptr);
+    for (size_t i = 0; i < out->size(); ++i) {
+      results[i] =
+          static_cast<phi::distributed::DistTensor*>(out->at(i).impl().get());
+    }
+  }
+  return results;
+}
+
 }  // namespace experimental
 }  // namespace paddle
@@ -150,5 +150,11 @@ std::vector<phi::distributed::DistTensor*> SetKernelDistOutput(
 std::vector<phi::distributed::DistTensor*> SetKernelDistOutput(
     size_t out_size, std::vector<Tensor>* out);
 
+std::vector<phi::distributed::DistTensor*> SetKernelDistInplaceOutput(
+    size_t out_size, std::vector<Tensor>* out);
+
+std::vector<phi::distributed::DistTensor*> SetKernelDistInplaceOptionalOutput(
+    size_t out_size, paddle::optional<std::vector<Tensor>> out);
+
 }  // namespace experimental
 }  // namespace paddle
@@ -672,7 +672,7 @@ PrepareDataForDistTensor(const std::vector<Tensor>& input,
                          const TransformFlag& transform_flag,
                          bool is_stride_kernel) {
   std::vector<std::shared_ptr<phi::distributed::DistTensor>> out;
-  for (auto x : input) {
+  for (auto& x : input) {
     const auto& tensor_in = x.impl();
     if (tensor_in) {
       phi::distributed::DistTensor* dist_tensor =
@@ -691,22 +691,46 @@ PrepareDataForDistTensor(const std::vector<Tensor>& input,
                                      dense_tensor.meta().is_contiguous()))) {
         out.push_back(
             std::static_pointer_cast<phi::distributed::DistTensor>(tensor_in));
-        continue;
+      } else {
+        phi::DenseTensor trans_in_tensor = TransformData(
+            dense_tensor, target_args_def, transform_flag, is_stride_kernel);
+        // TODO(GhostScreaming): The global meta in DistTensor is not changed,
+        // but the local meta in DenseTensor maybe changed, such as layout
+        // change(NCHW->NHWC), so the new DistTensor's meta maybe not unified.
+        VLOG(6) << "PrepareDataForDistTensor return transformed dist tensor";
+        out.push_back(std::make_shared<phi::distributed::DistTensor>(
+            trans_in_tensor, dist_tensor->dist_attr()));
       }
-      phi::DenseTensor trans_in_tensor = TransformData(
-          dense_tensor, target_args_def, transform_flag, is_stride_kernel);
-      // TODO(GhostScreaming): The global meta in DistTensor is not changed,
-      // but the local meta in DenseTensor maybe changed, such as layout
-      // change(NCHW->NHWC), so the new DistTensor's meta maybe not unified.
-      VLOG(6) << "PrepareDataForDistTensor return transformed dist tensor";
-      out.push_back(std::make_shared<phi::distributed::DistTensor>(
-          trans_in_tensor, dist_tensor->dist_attr()));
     } else {
       out.push_back(nullptr);
     }
   }
   return out;
 }
 
+paddle::optional<phi::distributed::DistTensor> PrepareDataForDistTensor(
+    const paddle::optional<Tensor>& input,
+    const phi::TensorArgDef& target_args_def,
+    const TransformFlag& transform_flag,
+    bool is_stride_kernel) {
+  if (input) {
+    return {*PrepareDataForDistTensor(
+        *input, target_args_def, transform_flag, is_stride_kernel)};
+  }
+  return paddle::none;
+}
+
+paddle::optional<std::vector<std::shared_ptr<phi::distributed::DistTensor>>>
+PrepareDataForDistTensor(const paddle::optional<std::vector<Tensor>>& input,
+                         const phi::TensorArgDef& target_args_def,
+                         const TransformFlag& transform_flag,
+                         bool is_stride_kernel) {
+  if (input) {
+    return PrepareDataForDistTensor(
+        *input, target_args_def, transform_flag, is_stride_kernel);
+  }
+  return paddle::none;
+}
+
 }  // namespace experimental
 }  // namespace paddle
@@ -198,5 +198,17 @@ PrepareDataForDistTensor(const std::vector<Tensor>& input,
                          const TransformFlag& transform_flag,
                          bool is_stride_kernel);
 
+paddle::optional<phi::distributed::DistTensor> PrepareDataForDistTensor(
+    const paddle::optional<Tensor>& input,
+    const phi::TensorArgDef& target_args_def,
+    const TransformFlag& transform_flag,
+    bool is_stride_kernel);
+
+paddle::optional<std::vector<std::shared_ptr<phi::distributed::DistTensor>>>
+PrepareDataForDistTensor(const paddle::optional<std::vector<Tensor>>& input,
+                         const phi::TensorArgDef& target_args_def,
+                         const TransformFlag& transform_flag,
+                         bool is_stride_kernel);
+
 }  // namespace experimental
 }  // namespace paddle