From 45073b7c3971771ab14d82baec09ca13b2fe3953 Mon Sep 17 00:00:00 2001
From: qingqing01 <dangqingqing@baidu.com>
Date: Thu, 15 Mar 2018 16:57:14 +0800
Subject: [PATCH] Always synchronize when copy data on GPU from C++ to Numpy
 array. (#9110)

---
 paddle/fluid/pybind/tensor_py.h | 1 +
 1 file changed, 1 insertion(+)

diff --git a/paddle/fluid/pybind/tensor_py.h b/paddle/fluid/pybind/tensor_py.h
index 3b206f2f8..6f8c597f8 100644
--- a/paddle/fluid/pybind/tensor_py.h
+++ b/paddle/fluid/pybind/tensor_py.h
@@ -72,6 +72,7 @@ struct CastToPyBufferImpl<true, I, ARGS...> {
         paddle::platform::GpuMemcpyAsync(
             dst_ptr, src_ptr, sizeof(CUR_TYPE) * tensor.numel(),
             cudaMemcpyDeviceToHost, dev_ctx->stream());
+        dev_ctx->Wait();
 #else
         PADDLE_THROW("'CUDAPlace' is not supported in CPU only device.");
 #endif
-- 
GitLab