diff --git a/src/c_api/c_api_ndarray.cc b/src/c_api/c_api_ndarray.cc
index 7efb31e225fd..c88548152dc8 100644
--- a/src/c_api/c_api_ndarray.cc
+++ b/src/c_api/c_api_ndarray.cc
@@ -393,7 +393,7 @@ void PushOperator(const OpStatePtr& state,
 #if MXNET_USE_CUDA
           CastNonDefaultStorage<gpu>(temp_in_src, temp_in_dst, opctx);
           fcompute(state, opctx, input_blobs, req, output_blobs);
-          CastNonDefaultStorage<gpu>(temp_our_dst, temp_out_src, opctx);
+          CastNonDefaultStorage<gpu>(temp_out_dst, temp_out_src, opctx);
 #else
           LOG(FATAL) << MXNET_GPU_NOT_ENABLED_ERROR;
 #endif
diff --git a/src/operator/tensor/sparse_retain.cu b/src/operator/tensor/sparse_retain.cu
new file mode 100644
index 000000000000..b41e986d44c5
--- /dev/null
+++ b/src/operator/tensor/sparse_retain.cu
@@ -0,0 +1,18 @@
+/*!
+ * Copyright (c) 2017 by Contributors
+ * \file sparse_retain.cu
+ * \brief
+*/
+
+#include "./sparse_retain-inl.h"
+namespace mxnet {
+namespace op {
+
+NNVM_REGISTER_OP(sparse_retain)
+.set_attr<FComputeEx>("FComputeEx<gpu>", SparseRetainOpForwardEx<gpu>);
+
+NNVM_REGISTER_OP(_backward_sparse_retain)
+.set_attr<FComputeEx>("FComputeEx<gpu>", SparseRetainOpBackwardEx<gpu>);
+
+}  // namespace op
+}  // namespace mxnet