"parallel op set lod after copy "

paddle/framework/mixed_vector.h

@@ -54,7 +54,7 @@ class Vector : public std::vector<T> {
     CopyToCUDA();
     PADDLE_ENFORCE_NOT_NULL(
         cuda_ptr_, "No data or Insufficient CUDA memory to allocation");
-    return static_cast<T *>(cuda_ptr_);
+    return static_cast<T *>(cuda_ptr_.get());
   }
 
   /* Get host vector */
@@ -127,7 +127,7 @@ void Vector<T>::CopyToCUDA() {
 #ifdef PADDLE_WITH_CUDA
   if (cuda_size_ < this->size() || cuda_ptr_ == nullptr) {
     cuda_ptr_.reset(
-        memory::Alloc<platform::CUDAPlace>(this->size() * sizeof(T)),
+        memory::Alloc<platform::CUDAPlace>(place_, this->size() * sizeof(T)),
         memory::PlainDeleter<void, platform::CUDAPlace>(place_));
   }
   cuda_size_ = this->size();

paddle/operators/parallel_do_op.cc

@@ -79,7 +79,7 @@ inline void CopyOrShare(const framework::Variable &src,
     } else {
       Copy(src.Get<LoDTensor>(), dst_place, dst->GetMutable<LoDTensor>());
     }
-    dst->set_lod(src.lod());
+    dst->GetMutable<LoDTensor>()->set_lod(src.Get<LoDTensor>().lod());
   } else if (src.IsType<SelectedRows>()) {
     auto &src_sr = src.Get<SelectedRows>();
     auto *dst_sr = dst->GetMutable<SelectedRows>();