diff --git a/src/operator/mshadow_op.h b/src/operator/mshadow_op.h
index bd9af2c148a9..8c2dc4708444 100644
--- a/src/operator/mshadow_op.h
+++ b/src/operator/mshadow_op.h
@@ -746,6 +746,10 @@ MXNET_BINARY_MATH_OP_NC(minus_sign, a - b > DType(0) ? DType(1) : -DType(1));
 
 MXNET_BINARY_MATH_OP(rminus, b - a);
 
+MXNET_BINARY_MATH_OP_NC(self_grad, 1);
+
+MXNET_BINARY_MATH_OP_NC(minus_rgrad, -1);
+
 MXNET_BINARY_MATH_OP(div_grad, 1.0f / math::id(b));
 
 MXNET_BINARY_MATH_OP(div_rgrad, -math::id(a) / math::sqr(b));
diff --git a/src/operator/numpy/np_elemwise_broadcast_op.cc b/src/operator/numpy/np_elemwise_broadcast_op.cc
index 41c9ea70eb61..48f6e97d35df 100644
--- a/src/operator/numpy/np_elemwise_broadcast_op.cc
+++ b/src/operator/numpy/np_elemwise_broadcast_op.cc
@@ -115,7 +115,22 @@ MXNET_OPERATOR_REGISTER_NP_BINARY_MIXED_PRECISION(_npi_add)
   "FCompute<cpu>",
   NumpyBinaryBroadcastComputeWithBool<cpu, op::mshadow_op::plus>)
 #endif
-.set_attr<nnvm::FGradient>("FGradient", ElemwiseGradUseNone{"_backward_broadcast_add"});
+.set_attr<nnvm::FGradient>("FGradient", ElemwiseGradUseIn{"_backward_npi_broadcast_add"});
+
+NNVM_REGISTER_OP(_backward_npi_broadcast_add)
+.set_num_inputs(3)
+.set_num_outputs(2)
+.set_attr<nnvm::TIsBackward>("TIsBackward", true)
+.set_attr<nnvm::FInplaceOption>("FInplaceOption",
+  [](const NodeAttrs& attrs){
+    return std::vector<std::pair<int, int> >{{0, 1}};
+  })
+.set_attr<FResourceRequest>("FResourceRequest",
+  [](const NodeAttrs& attrs) {
+    return std::vector<ResourceRequest>{ResourceRequest::kTempSpace};
+  })
+.set_attr<FCompute>("FCompute<cpu>", NumpyBinaryBackwardUseIn<cpu, mshadow_op::self_grad,
+                                                                mshadow_op::self_grad>);
 
 MXNET_OPERATOR_REGISTER_NP_BINARY_MIXED_PRECISION(_npi_subtract)
 #ifndef _WIN32
@@ -128,7 +143,22 @@ MXNET_OPERATOR_REGISTER_NP_BINARY_MIXED_PRECISION(_npi_subtract)
   "FCompute<cpu>",
   NumpyBinaryBroadcastCompute<cpu, op::mshadow_op::minus>)
 #endif
-.set_attr<nnvm::FGradient>("FGradient", ElemwiseGradUseNone{"_backward_broadcast_sub"});
+.set_attr<nnvm::FGradient>("FGradient", ElemwiseGradUseIn{"_backward_npi_broadcast_sub"});
+
+NNVM_REGISTER_OP(_backward_npi_broadcast_sub)
+.set_num_inputs(3)
+.set_num_outputs(2)
+.set_attr<nnvm::TIsBackward>("TIsBackward", true)
+.set_attr<nnvm::FInplaceOption>("FInplaceOption",
+  [](const NodeAttrs& attrs){
+    return std::vector<std::pair<int, int> >{{0, 1}};
+  })
+.set_attr<FResourceRequest>("FResourceRequest",
+  [](const NodeAttrs& attrs) {
+    return std::vector<ResourceRequest>{ResourceRequest::kTempSpace};
+  })
+.set_attr<FCompute>("FCompute<cpu>", NumpyBinaryBackwardUseIn<cpu, mshadow_op::self_grad,
+                                                                mshadow_op::minus_rgrad>);
 
 MXNET_OPERATOR_REGISTER_NP_BINARY_MIXED_PRECISION(_npi_multiply)
 #ifndef _WIN32
diff --git a/src/operator/numpy/np_true_divide-inl.h b/src/operator/numpy/np_true_divide-inl.h
index 0bc60a08803e..c0edc8cad47a 100644
--- a/src/operator/numpy/np_true_divide-inl.h
+++ b/src/operator/numpy/np_true_divide-inl.h
@@ -29,6 +29,7 @@
 #include <vector>
 #include "../../common/utils.h"
 #include "../tensor/elemwise_binary_broadcast_op.h"
+#include "../numpy/np_elemwise_broadcast_op.h"
 
 namespace mxnet {
 namespace op {
diff --git a/src/operator/numpy/np_true_divide.cc b/src/operator/numpy/np_true_divide.cc
index 1e46cc9d13b5..bee8a80bab7f 100644
--- a/src/operator/numpy/np_true_divide.cc
+++ b/src/operator/numpy/np_true_divide.cc
@@ -80,10 +80,26 @@ NNVM_REGISTER_OP(_npi_true_divide)
   })
 #endif
 .set_attr<FCompute>("FCompute<cpu>", TrueDivideBroadcastCompute<cpu>)
-.set_attr<nnvm::FGradient>("FGradient", ElemwiseGradUseIn{"_backward_broadcast_div"})
+.set_attr<nnvm::FGradient>("FGradient", ElemwiseGradUseIn{"_backward_npi_broadcast_div"})
 .add_argument("lhs", "NDArray-or-Symbol", "Dividend array")
 .add_argument("rhs", "NDArray-or-Symbol", "Divisor array");
 
+
+NNVM_REGISTER_OP(_backward_npi_broadcast_div)
+.set_num_inputs(3)
+.set_num_outputs(2)
+.set_attr<nnvm::TIsBackward>("TIsBackward", true)
+.set_attr<nnvm::FInplaceOption>("FInplaceOption",
+  [](const NodeAttrs& attrs){
+    return std::vector<std::pair<int, int> >{{0, 1}};
+  })
+.set_attr<FResourceRequest>("FResourceRequest",
+  [](const NodeAttrs& attrs) {
+    return std::vector<ResourceRequest>{ResourceRequest::kTempSpace};
+  })
+.set_attr<FCompute>("FCompute<cpu>", NumpyBinaryBackwardUseIn<cpu, mshadow_op::div_grad,
+                                                              mshadow_op::div_rgrad>);
+
 NNVM_REGISTER_OP(_npi_true_divide_scalar)
 .set_num_inputs(1)
 .set_num_outputs(1)
diff --git a/src/operator/operator_tune.cc b/src/operator/operator_tune.cc
index 4077ce137463..211b90f2396a 100644
--- a/src/operator/operator_tune.cc
+++ b/src/operator/operator_tune.cc
@@ -430,6 +430,8 @@ IMPLEMENT_BINARY_WORKLOAD_FWD(mxnet::op::mshadow_op::rldexp);  // NOLINT()
 IMPLEMENT_BINARY_WORKLOAD_BWD(mxnet::op::mshadow_op::ldexp_grad);  // NOLINT()
 IMPLEMENT_BINARY_WORKLOAD_BWD(mxnet::op::mshadow_op::ldexp_rgrad);  // NOLINT()
 IMPLEMENT_BINARY_WORKLOAD_BWD(mxnet::op::mshadow_op::rldexp_grad);  // NOLINT()
+IMPLEMENT_BINARY_WORKLOAD_BWD(mxnet::op::mshadow_op::self_grad); // NOLINT()
+IMPLEMENT_BINARY_WORKLOAD_BWD(mxnet::op::mshadow_op::minus_rgrad); // NOLINT()
 /*!
  * \brief Tuner objects, *not* automatically generated
  */
diff --git a/tests/python/unittest/test_numpy_op.py b/tests/python/unittest/test_numpy_op.py
index 712a846a556b..8b419160e0b6 100644
--- a/tests/python/unittest/test_numpy_op.py
+++ b/tests/python/unittest/test_numpy_op.py
@@ -2584,8 +2584,10 @@ def hybrid_forward(self, F, a, b, *args, **kwargs):
                             use_broadcast=False, equal_nan=True)
 
     funcs = {
-        'add': (-1.0, 1.0, None, None),
-        'subtract': (-1.0, 1.0, None, None),
+        'add': (-1.0, 1.0, lambda y, x1, x2: _np.ones(y.shape),
+                           lambda y, x1, x2: _np.ones(y.shape)),
+        'subtract': (-1.0, 1.0, lambda y, x1, x2: _np.ones(y.shape),
+                                lambda y, x1, x2: _np.ones(y.shape) * -1),
         'multiply': (-1.0, 1.0, lambda y, x1, x2: _np.broadcast_to(x2, y.shape),
                                 lambda y, x1, x2: _np.broadcast_to(x1, y.shape)),
         'power': (1.0, 3.0, lambda y, x1, x2: _np.power(x1, x2 - 1.0) * x2,