"FusedBatchNorm requires at least 5 outputs");
  }
  if (grad_inputs.empty()) {
    return errors::InvalidArgument("FusedBatchNorm grad requires 1 grad input");
  }
  if (op.num_inputs() < 3) {
    return errors::InvalidArgument("FusedBatchNorm has too few inputs");
  }

  Output x = op.input(0);
  Output grad_y = grad_inputs[0];
  Output scale = op.input(1);
  float epsilon;

/// @addtogroup core
/// @{

/// Represents a node in the computation graph.
class Operation {
 public:
  Operation() : node_(nullptr) {}
  explicit Operation(Node* n);

  int32 num_inputs() const { return node_->num_inputs(); }
  DataType input_type(int32_t o) const { return node_->input_type(o); }
  Output input(int32_t i) const;

  int32 num_outputs() const { return node_->num_outputs(); }

  // hence dx_k = dy for all x_k
  // So the gradient for AddN just transfers the incoming gradient to
  // all outgoing gradients.
  auto incoming = Identity(scope, grad_inputs[0]);
  for (int32_t i = 0; i < op.num_inputs(); ++i) {
    grad_outputs->push_back(incoming);
  }
  return scope.status();
}
REGISTER_GRADIENT_OP("AddN", AddNGrad);

Status PowGrad(const Scope& scope, const Operation& op,

      reinterpret_cast<NamedParallelDevice*>(device_info);
  std::vector<MaybeParallelTensorUnowned> typed_inputs;
  int num_inputs = TFE_OpGetFlatInputCount(original_op, status);
  if (TF_GetCode(status) != TF_OK) return;
  typed_inputs.reserve(num_inputs);
  for (int i = 0; i < num_inputs; ++i) {
    TFE_TensorHandle* input = TFE_OpGetFlatInput(original_op, i, status);
    if (TF_GetCode(status) != TF_OK) return;

  ASSERT_TRUE(status.ok()) << status.ToString();

  EXPECT_EQ("XTimesY", kernel_->name());
  EXPECT_EQ("XTimesY", kernel_->type_string());

  EXPECT_EQ(2, kernel_->num_inputs());
  EXPECT_EQ(DT_FLOAT, kernel_->input_type(0));
  EXPECT_EQ(DT_RESOURCE, kernel_->input_type(1));
  EXPECT_EQ(DEVICE_MEMORY, kernel_->input_memory_types()[0]);
  EXPECT_EQ(HOST_MEMORY, kernel_->input_memory_types()[1]);

  }
#endif  // defined(IS_MOBILE_PLATFORM) || defined(IS_SLIM_BUILD)
}

int TF_NumInputs(TF_OpKernelContext* ctx) {
  auto* cc_ctx = reinterpret_cast<::tensorflow::OpKernelContext*>(ctx);
  return cc_ctx->num_inputs();
}

int TF_NumOutputs(TF_OpKernelContext* ctx) {
  auto* cc_ctx = reinterpret_cast<::tensorflow::OpKernelContext*>(ctx);
  return cc_ctx->num_outputs();
}

  // grad = [[g_1, g_2], [g_3, g_4], [g_5, g_6]]

  // indices and data are two equal-sized lists passed
  // into DynamicStitch.
  // num_values = 2
  int32_t num_values = op.num_inputs() / 2;

  // Stop propagation along the indices list
  for (int32_t i = 0; i < num_values; i++) {
    grad_outputs->push_back(NoGradient());
  }

// returned through the provided TF_OutputList.
// Any active tape will observe the effects of this execution.
void TF_ExecuteOperation(TF_AbstractOp* op, int num_inputs,
                         TF_AbstractTensor* const* inputs, TF_OutputList* o,
                         TF_Status* s);

// Creates a new TF_AbstractFunction from the current tracing states in the

      .Attr("SomeDataTypeAttr: type");

  static int num_inputs = 0;
  static int num_outputs = 0;

  // A kernel whose Compute function has a side-effect of updating num_inputs
  // and num_outputs. Various functions on TF_OpKernelContext are also
  // exercised.
  auto my_compute_func = [](void* kernel, TF_OpKernelContext* ctx) {
    num_inputs = TF_NumInputs(ctx);
    num_outputs = TF_NumOutputs(ctx);

        absl::AbortedError(
            "The `ids` and `calibration_methods` must have the same size."));

    // Check the number and type of inputs.
    OP_REQUIRES(context, context->num_inputs() == ids_.size() * 3,
                absl::AbortedError("The number of inputs must be  three times "
                                   "the size of the `ids` list."));
    for (int i = 0; i < ids_.size(); ++i) {