namespace gradients {
namespace internal {

void CompareNumericalAndAutodiffGradients(
    Model model, Model grad_model, AbstractContext* ctx,
    absl::Span<AbstractTensorHandle* const> inputs, bool use_function,
    double abs_error) {
  auto num_inputs = inputs.size();
  std::vector<AbstractTensorHandle*> outputs(num_inputs);
  auto s = RunModel(grad_model, ctx, inputs, absl::MakeSpan(outputs),
                    /*use_function=*/use_function);

      RunModel(forward, ctx, inputs, model_outputs, use_function));
  AbstractTensorHandlePtr model_out(model_outputs[0]);

  TF_Tensor* model_out_tensor;
  TF_RETURN_IF_ERROR(GetValue(model_out.get(), &model_out_tensor));
  int num_dims_out = TF_NumDims(model_out_tensor);
  TF_DeleteTensor(model_out_tensor);

  // If the output is a scalar, then return the scalar output
  if (num_dims_out == 0) {
    outputs[0] = model_out.release();

      TF_RETURN_IF_ERROR(
          CreateParamsForInputs(func_ctx.get(), inputs, &func_inputs));
      std::vector<AbstractTensorHandle*> model_outputs;
      model_outputs.resize(outputs.size());
      TF_RETURN_IF_ERROR(model(func_ctx.get(), absl::MakeSpan(func_inputs),
                               absl::MakeSpan(model_outputs)));
      for (auto func_input : func_inputs) {
        func_input->Unref();
      }
      AbstractFunction* func = nullptr;

  TF_RETURN_IF_ERROR(ops::SparseSoftmaxCrossEntropyWithLogits(
      ctx, inputs[0], inputs[1], &loss, &backprop,
      "SparseSoftmaxCrossEntropyWithLogits"));
  // `gradient_checker` only works with model that returns only 1 tensor.
  // Although, `ops::SparseSoftmaxCrossEntropyWithLogits` returns 2 tensors, the
  // second tensor isn't needed for computing gradient so we could safely drop
  // it.
  outputs[0] = loss;