Add wait

paddle/fluid/pybind/tensor_py.h

@@ -190,6 +190,7 @@ void PyCUDATensorSetFromArray(
       static_cast<const platform::CUDADeviceContext *>(pool.Get(place));
   paddle::platform::GpuMemcpyAsync(dst, array.data(), sizeof(T) * array.size(),
                                    cudaMemcpyHostToDevice, dev_ctx->stream());
+  dev_ctx->Wait();
 }
 
 template <>