[PHI] Unify expand_as, expand_as_grad kernel behavior between cpu and gpu (#73101)

ooooo-create · web-flow · commit 23748bf8351d · 2025-06-13T10:05:32.000+08:00
* unify behavior

* refine code
diff --git a/paddle/phi/kernels/impl/expand_as_grad_kernel_impl.h b/paddle/phi/kernels/impl/expand_as_grad_kernel_impl.h
@@ -49,18 +49,21 @@ void ExpandAsGradKernel(const Context& context,
                         const std::vector<int64_t>& target_shape,
                         DenseTensor* in_grad) {
   auto x_dims = x.dims();
+  auto out_grad_dims = out_grad.dims();
+  std::vector<int64_t> real_target_shape =
+      phi::vectorize<int64_t>(out_grad_dims);
 
-  if (in_grad->dims() == out_grad.dims()) {
+  if (in_grad->dims() == out_grad_dims) {
     phi::Copy(context, out_grad, context.GetPlace(), false, in_grad);
     return;
   }
 
   auto vec_in_dims = common::vectorize<int64_t>(x_dims);
-  auto diff = target_shape.size() - vec_in_dims.size();
+  auto diff = real_target_shape.size() - vec_in_dims.size();
   vec_in_dims.insert(vec_in_dims.begin(), diff, 1);
   std::vector<int64_t> repeat_times(vec_in_dims.size());
   for (size_t i = 0; i < vec_in_dims.size(); ++i) {
-    repeat_times[i] = target_shape[i] / vec_in_dims[i];
+    repeat_times[i] = real_target_shape[i] / vec_in_dims[i];
   }
   std::vector<int64_t> reshape_dims_vec;
   std::vector<int> reduce_dims_vec;
diff --git a/paddle/phi/kernels/impl/expand_as_kernel_impl.h b/paddle/phi/kernels/impl/expand_as_kernel_impl.h
@@ -125,15 +125,8 @@ void ExpandAsKernel(const Context& dev_ctx,
                         MAX_RANK_SUPPORTED));
 
   std::vector<int64_t> real_target_shape = target_shape;
-  for (size_t i = 0; i < target_shape.size(); ++i) {
-    if (target_shape[i] == -1) {
-      if (y) {
-        if (y->IsInitialized()) {
-          real_target_shape = common::vectorize<int64_t>(y->dims());
-        }
-      }
-      break;
-    }
+  if (y.get_ptr()) {
+    real_target_shape = phi::vectorize<int64_t>(y.get_ptr()->dims());
   }
 
   switch (target_rank) {
