absl::Span<AbstractTensorHandle* const> grad_outputs,
                 absl::Span<AbstractTensorHandle*> grad_inputs) override {
    // Grad for Softmax Input
    TF_RETURN_IF_ERROR(BroadcastMul(
        ctx, grad_outputs[0], forward_outputs_[1],
        grad_inputs.subspan(0, 1)));  // upstream_grad * local softmax grad

    // Grad for labels is null
    grad_inputs[1] = nullptr;
    return absl::OkStatus();
  }

limitations under the License.
==============================================================================*/
#ifndef TENSORFLOW_C_EXPERIMENTAL_GRADIENTS_NN_GRAD_H_
#define TENSORFLOW_C_EXPERIMENTAL_GRADIENTS_NN_GRAD_H_

#include "tensorflow/c/eager/gradients.h"

namespace tensorflow {
namespace gradients {
GradientFunction* ReluRegisterer(const ForwardOperation& op);

#ifndef TENSORFLOW_C_EXPERIMENTAL_GRADIENTS_GRAD_TEST_HELPER_H_
#define TENSORFLOW_C_EXPERIMENTAL_GRADIENTS_GRAD_TEST_HELPER_H_

#include "tensorflow/c/eager/gradients.h"
#include "tensorflow/c/eager/unified_api_testutil.h"

namespace tensorflow {
namespace gradients {
namespace internal {

void CompareNumericalAndAutodiffGradients(
    Model model, Model grad_model, AbstractContext* ctx,

            n->name(),
            "). This is a bug. Please file an issue at "
            "https://github.com/tensorflow/tensorflow/issues.");
      }
    }
  }

  // Unpack the results from grad_outputs_arg.
  TFOutputsFromOutputs(dy_arg, dy);
#endif  // defined(IS_MOBILE_PLATFORM) || defined(IS_SLIM_BUILD)
}

// TF_Session functions ----------------------------------------------

    TF_Tensor* grad_tensor;
    TF_RETURN_IF_ERROR(GetValue(diff_quotient.get(), &grad_tensor));
    float grad_data[1];
    memcpy(&grad_data[0], TF_TensorData(grad_tensor),
           TF_TensorByteSize(grad_tensor));
    TF_DeleteTensor(grad_tensor);
    dtheta_approx[i] = grad_data[0];
  }

  // Populate *numerical_grad with the data from dtheta_approx.

namespace tensorflow {
namespace gradients {
// Ignores `grad_outputs` and sets all entries in grad_inputs to nullptr.
class NotDifferentiableGradientFunction : public GradientFunction {
  Status Compute(AbstractContext* ctx,
                 absl::Span<AbstractTensorHandle* const> grad_outputs,
                 absl::Span<AbstractTensorHandle*> grad_inputs) override;
};

  if (TF_GetCode(status) != TF_OK) return;
  if (!grad) return;

  status->status = g->graph.AddFunctionDef(grad->record->fdef(),
                                           grad->record->stack_traces());
  if (TF_GetCode(status) != TF_OK) return;

  tensorflow::GradientDef gdef;
  gdef.set_function_name(func->record->fdef().signature().name());
  gdef.set_gradient_func(grad->record->fdef().signature().name());

#include "tensorflow/c/eager/gradients_internal.h"
#include "tensorflow/c/eager/unified_api_testutil.h"
#include "tensorflow/c/experimental/gradients/array_grad.h"
#include "tensorflow/c/experimental/gradients/math_grad.h"
#include "tensorflow/c/experimental/gradients/not_differentiable.h"
#include "tensorflow/c/experimental/gradients/tape/tape_context.h"
#include "tensorflow/c/experimental/ops/array_ops.h"

// Adds a copy of function `func` and optionally its gradient function `grad`
// to `g`. Once `func`/`grad` is added to `g`, it can be called by creating
// an operation using the function's name.
// Any changes to `func`/`grad` (including deleting it) done after this method
// returns, won't affect the copy of `func`/`grad` in `g`.
// If `func` or `grad` are already in `g`, TF_GraphCopyFunction has no

  TFE_OpSetAttrType(op.get(), "dtype", type);
  TFE_OpSetAttrShape(op.get(), "shape", dims, num_dims, status);
  TFE_OpSetAttrString(op.get(), "container", "", 0);
  // Use the special GUID for no buffer sharing
  //
  // TODO(allenl): Should we provide a better API for this? AFAIK this is the
  // only reasonable way to make variables with no aliasing using the eager C
  // API.