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();
  }

            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 ----------------------------------------------

      auto grad_input = grad_outputs[i];
      // TODO(srbs): Should we add a copy contructor to AbstractTensorHandle
      // that takes care of this similar to `Tensor`?
      if (grad_input) {
        grad_input->Ref();
      }
      grad_inputs[i] = grad_input;
    }
    return absl::OkStatus();
  }
  ~IdentityNGradientFunction() override {}
};
}  // namespace

                            &numerical_grad_raw);
      ASSERT_EQ(errors::OK, s.code()) << s.message();
      numerical_grad.reset(numerical_grad_raw);
    }

    TF_Tensor* numerical_tensor;
    s = GetValue(numerical_grad.get(), &numerical_tensor);
    ASSERT_EQ(errors::OK, s.code()) << s.message();
    auto num_elem_numerical = TF_TensorElementCount(numerical_tensor);

==============================================================================*/
#include "tensorflow/c/experimental/gradients/math_grad.h"

#include "tensorflow/c/eager/c_api_test_util.h"
#include "tensorflow/c/eager/c_api_unified_experimental_internal.h"
#include "tensorflow/c/eager/unified_api_testutil.h"
#include "tensorflow/c/experimental/gradients/grad_test_helper.h"
#include "tensorflow/c/experimental/gradients/tape/tape_context.h"

    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.

  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"

 public:
  Status Compute(AbstractContext* ctx,
                 absl::Span<AbstractTensorHandle* const> grad_outputs,
                 absl::Span<AbstractTensorHandle*> grad_inputs) override {
    CHECK_EQ(grad_outputs.size(), 1);
    CHECK_EQ(grad_inputs.size(), 1);
    grad_inputs[0] = grad_outputs[0];
    if (grad_inputs[0]) {
      grad_inputs[0]->Ref();
    }
    return absl::OkStatus();
  }
};

// Computes:
//

  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.