diff --git a/paddle/fluid/framework/fleet/heter_ps/graph_gpu_ps_table_inl.h b/paddle/fluid/framework/fleet/heter_ps/graph_gpu_ps_table_inl.h
index 605019cb607fc41bd32bd9053128fd4791bb2c40..d28ae0ab5d93f484b8ebdfca4099d63eab080812 100644
--- a/paddle/fluid/framework/fleet/heter_ps/graph_gpu_ps_table_inl.h
+++ b/paddle/fluid/framework/fleet/heter_ps/graph_gpu_ps_table_inl.h
@@ -924,9 +924,11 @@ NeighborSampleResult GpuPsGraphTable::graph_neighbor_sample_v2(
   {
     platform::CUDAPlace place = platform::CUDAPlace(resource_->dev_id(gpu_id));
     platform::CUDADeviceGuard guard(resource_->dev_id(gpu_id));
-    thrust::device_ptr<int> t_actual_sample_size(actual_sample_size);
-    int total_sample_size =
-        thrust::reduce(t_actual_sample_size, t_actual_sample_size + len);
+    thrust::device_vector<int> t_actual_sample_size(len);
+    thrust::copy(actual_sample_size, actual_sample_size + len,
+                 t_actual_sample_size.begin());
+    int total_sample_size = thrust::reduce(t_actual_sample_size.begin(),
+                                           t_actual_sample_size.end());
     result.actual_val_mem =
         memory::AllocShared(place, total_sample_size * sizeof(int64_t));
     result.actual_val = (int64_t*)(result.actual_val_mem)->ptr();
@@ -934,7 +936,8 @@ NeighborSampleResult GpuPsGraphTable::graph_neighbor_sample_v2(
     result.set_total_sample_size(total_sample_size);
 
     thrust::device_vector<int> cumsum_actual_sample_size(len);
-    thrust::exclusive_scan(t_actual_sample_size, t_actual_sample_size + len,
+    thrust::exclusive_scan(t_actual_sample_size.begin(),
+                           t_actual_sample_size.end(),
                            cumsum_actual_sample_size.begin(), 0);
     fill_actual_vals<<<grid_size, block_size_, 0, stream>>>(
         val, result.actual_val, actual_sample_size,