/*sources=*/inputs,
                                          /*output_gradients=*/{},
                                          /*result=*/outputs));
  exp_output->Unref();
  return absl::OkStatus();
}

TEST_P(CustomGradientTest, ExpWithPassThroughGrad) {
  std::unique_ptr<TF_Status, decltype(&TF_DeleteStatus)> status(
      TF_NewStatus(), TF_DeleteStatus);
  AbstractContextPtr ctx;

      ASSERT_NEAR(dnumerical[j], danalytical[j], abs_error);
    }
    TF_DeleteTensor(analytical_tensor);
    TF_DeleteTensor(numerical_tensor);
    delete[] danalytical;
    delete[] dnumerical;
    outputs[i]->Unref();
  }
}

void CheckTensorValue(AbstractTensorHandle* t, absl::Span<const float> manuals,
                      absl::Span<const int64_t> dims, double abs_error) {
  TF_Tensor* analytical_tensor;

                                &grad_inputs[0], name.c_str()));
    return absl::OkStatus();
  }
  ~ReluGradientFunction() override {
    for (auto output : forward_outputs_) {
      if (output) {
        output->Unref();
      }
    }
  }

 private:
  // TODO(b/174778737): Only hold needed outputs.
  vector<AbstractTensorHandle*> forward_outputs_;
};

Status BroadcastMul(AbstractContext* ctx, AbstractTensorHandle* vec,

  // Although, `ops::IdentityN` returns 2 tensors, the first tensor isn't needed
  // for computing gradient so we could safely drop it.
  outputs[0] = temp_outputs[1];
  temp_outputs[0]->Unref();
  return absl::OkStatus();
}

class CppGradients
    : public ::testing::TestWithParam<std::tuple<const char*, bool, bool>> {
 protected:
  void SetUp() override {
    TF_StatusPtr status(TF_NewStatus());

      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;
      OutputList output_list;
      output_list.expected_num_outputs = 0;
      output_list.outputs.reserve(outputs.size());

  // Although, `ops::SparseSoftmaxCrossEntropyWithLogits` returns 2 tensors, the
  // second tensor isn't needed for computing gradient so we could safely drop
  // it.
  outputs[0] = loss;
  backprop->Unref();
  return absl::OkStatus();
}

Status BiasAddModel(AbstractContext* ctx,
                    absl::Span<AbstractTensorHandle* const> inputs,
                    absl::Span<AbstractTensorHandle*> outputs) {