AbstractTensorHandle* expand_dims_outputs;
  TF_RETURN_IF_ERROR(
      ops::ExpandDims(ctx, vec, dim.get(), &expand_dims_outputs, "ExpandDims"));
  TF_RETURN_IF_ERROR(
      ops::Mul(ctx, expand_dims_outputs, mat, &outputs[0], "Mul"));
  expand_dims_outputs->Unref();
  return absl::OkStatus();
}

class SparseSoftmaxCrossEntropyWithLogitsGradientFunction
    : public GradientFunction {
 public:

  TF_SetAttrInt(reduce_desc, "group_size", group_size);
  TF_SetAttrInt(reduce_desc, "group_key", 0);
  TF_SetAttrInt(reduce_desc, "instance_key", 0);

  const std::string merge_op("Mul");
  TF_SetAttrString(reduce_desc, "merge_op", merge_op.c_str(),
                   merge_op.length());
  const std::string final_op("Id");
  TF_SetAttrString(reduce_desc, "final_op", final_op.c_str(),

TensorHandlePtr Multiply(TFE_Context* context, TFE_TensorHandle* first,
                         TFE_TensorHandle* second, TF_Status* status) {
  std::unique_ptr<TFE_Op, decltype(&TFE_DeleteOp)> op(
      TFE_NewOp(context, "Mul", status), TFE_DeleteOp);
  if (TF_GetCode(status) != TF_OK) return nullptr;
  TFE_OpAddInput(op.get(), first, status);
  if (TF_GetCode(status) != TF_OK) return nullptr;
  TFE_OpAddInput(op.get(), second, status);

}

Status MulModel(AbstractContext* ctx,
                absl::Span<AbstractTensorHandle* const> inputs,
                absl::Span<AbstractTensorHandle*> outputs) {
  return ops::Mul(ctx, inputs[0], inputs[1], &outputs[0], "Mul");
}

Status Log1pModel(AbstractContext* ctx,
                  absl::Span<AbstractTensorHandle* const> inputs,
                  absl::Span<AbstractTensorHandle*> outputs) {

    /* Given upstream grad U and a mul op A*B, the gradients are:
     *
     *    dA = U * B
     *    dB = A * U
     *
     */

    AbstractTensorHandle* upstream_grad = grad_outputs[0];

    // Gradient for A
    std::string name = "Mul_Grad_A";
    TF_RETURN_IF_ERROR(Mul(ctx, upstream_grad, forward_inputs_[1],

  EXPECT_EQ(2, product[0]);
  EXPECT_EQ(4, product[1]);
  EXPECT_EQ(6, product[2]);
  EXPECT_EQ(8, product[3]);

  // When we turn on grappler, confirm that the tf.Add has been rewritten into a
  // tf.Mul.
  // This capability of checking the executed op names is currently only enabled
  // for TFRT debug build, for performance and simplicity reasons.
  if (use_tfrt) {
    TF_Buffer* buf = TF_NewBuffer();

}

Status MulModel(AbstractContext* ctx,
                absl::Span<AbstractTensorHandle* const> inputs,
                absl::Span<AbstractTensorHandle*> outputs) {
  return ops::Mul(ctx, inputs[0], inputs[1], &outputs[0], "Mul");
}

// TODO(vnvo2409): Add more tests from `python/ops/gradient_checker_v2_test.py`.
// These tests should not be confused with `[*]_grad_test` which compare the

TF_Operation* Min(TF_Operation* l, TF_Operation* r, TF_Graph* graph,
                  TF_Status* s, const char* name = "min");

TF_Operation* Mul(TF_Operation* l, TF_Operation* r, TF_Graph* graph,
                  TF_Status* s, const char* name = "mul");

// If `op_device` is non-empty, set the created op on that device.
TF_Operation* MinWithDevice(TF_Operation* l, TF_Operation* r, TF_Graph* graph,

TEST_F(CApiGradientsTest, MultipleCallsToAddGradients) {
  const float X = 3.0f, Y = 7.0f;
  TF_Operation* x = Placeholder(graph_, s_, "x", TF_FLOAT);
  TF_Operation* y = Placeholder(graph_, s_, "y", TF_FLOAT);
  TF_Operation* xy = Mul(x, y, graph_, s_, "xy");
  TF_Output dxy_dx, dxy_dy;

  TF_Output outputs[1] = {{xy, 0}};
  TF_Output inputs[1] = {{x, 0}};
  TF_AddGradients(graph_, outputs, 1, inputs, 1, nullptr, s_, &dxy_dx);

                  TF_Status* s, const char* name) {
  return MinWithDevice(l, r, graph, /*op_device=*/"", s, name);
}

TF_Operation* Mul(TF_Operation* l, TF_Operation* r, TF_Graph* graph,
                  TF_Status* s, const char* name) {
  TF_Operation* op;
  BinaryOpHelper("Mul", l, r, graph, s, name, &op, "", true);
  return op;
}

TF_Operation* Add(TF_Output l, TF_Output r, TF_Graph* graph, TF_Status* s,