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();
  }

                 absl::Span<AbstractTensorHandle* const> grad_outputs,
                 absl::Span<AbstractTensorHandle*> grad_inputs) override {
    // TODO(b/161805092): Support broadcasting.

    DCHECK(grad_outputs[0]);
    grad_inputs[0] = grad_outputs[0];
    grad_inputs[1] = grad_outputs[0];

    grad_inputs[0]->Ref();
    grad_inputs[1]->Ref();
    return absl::OkStatus();
  }

      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/array_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"

limitations under the License.
==============================================================================*/
#include "tensorflow/c/experimental/gradients/nn_grad.h"

#include "tensorflow/c/eager/c_api_test_util.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"
#include "tensorflow/c/experimental/ops/nn_ops.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.

    ],
)

cc_library(
    name = "gradients",
    hdrs = [
        "array_grad.h",
        "math_grad.h",
        "nn_grad.h",
        "not_differentiable.h",
    ],
    visibility = [
        "//tensorflow:internal",
    ],
    deps = [
        ":array_grad",
        ":math_grad",
        ":nn_grad",
        ":not_differentiable",
        "//tensorflow/c/eager:abstract_context",

    TF_DeleteStatus(s_);
  }

  void TestGradientsSuccess(bool grad_inputs_provided) {
    TF_Output inputs[2];
    TF_Output outputs[1];
    TF_Output grad_outputs[2];
    TF_Output expected_grad_outputs[2];

    BuildSuccessGraph(inputs, outputs);
    BuildExpectedGraph(grad_inputs_provided, expected_grad_outputs);

    AddGradients(grad_inputs_provided, nullptr, inputs, 2, outputs, 1,

    int64_t tensor_id = source_tensor_ids[i];
    auto grad_it = gradients.find(tensor_id);
    if (grad_it == gradients.end()) {
      result[i] = nullptr;
    } else {
      if (grad_it->second.size() > 1) {
        Gradient* grad = vspace.AggregateGradients(grad_it->second);
        grad_it->second.clear();
        grad_it->second.push_back(grad);
      }
      result[i] = grad_it->second[0];
      used_gradient_ids.insert(tensor_id);