const char* attr_name, const void* proto,
                                size_t proto_len, TF_Status* status) {
  // shape.ParseFromArray takes an int as length, this function takes size_t,
  // make sure there is no information loss.
  if (proto_len > std::numeric_limits<int>::max()) {
    status->status = InvalidArgument(
        "proto_len (", proto_len,

    for (int i = 0; i < grad_outputs.size(); i++) {
      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 {}

      ParallelTensor::FromTensorHandles(parallel_device,
                                        std::move(mixed_handles), status.get());
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK);
  // Can't take the shape of a parallel tensor with varying numbers of axes, but
  // running operations on them is OK.
  TF_ASSERT_OK(unknown_length_vector->Shape(&shape));
  EXPECT_THAT(*shape, ElementsAre(-1));

  std::vector<Tensor> all_input_tensors;
  // Update the vector with information from `input_tensors` if provided.
  if (input_tensors != nullptr) {
    // Note that we take the address of the elements in `all_input_tensors`
    // below. Allocate enough space so that no reallocation happens, which will
    // make the pointers invalid.
    all_input_tensors.reserve(num_inputs);

        //
        // TODO(allenl): There may be smarter ways to do this copy in some
        // cases, i.e. with a collective broadcast. We'll need to be careful
        // about things that are taken as inputs on the host or on their
        // existing device (for multi-device functions).
        std::unique_ptr<ParallelTensor> parallel_tensor(
            parallel_device.CopyToParallelDevice(

    theta_inputs[input_index] = thetaMinus.get();
    TF_RETURN_IF_ERROR(
        RunAndMaybeSum(ctx, forward, theta_inputs, f_outputs, use_function));
    AbstractTensorHandlePtr fMinus(f_outputs[0]);

    // Take Difference of both estimates: (f(theta + eps) - f(theta - eps)).
    TF_RETURN_IF_ERROR(
        ops::Sub(ctx, fPlus.get(), fMinus.get(), f_outputs, "sub_top"));
    AbstractTensorHandlePtr fDiff(f_outputs[0]);

    }
    return tensorflow::errors::Internal("Graph pass runs for more than once!");
  }

 private:
  bool first_call_ = true;
};

// After the graph pass is registered, it takes effect globally and can affect
// other test cases. Define a static variable to switch it on and off.
bool GraphErrorInjectionPass::enabled_ = false;

}

bool ModularFileSystem::FilesExist(const std::vector<std::string>& files,
                                   TransactionToken* token,
                                   std::vector<Status>* status) {
  if (ops_->paths_exist == nullptr)
    return FileSystem::FilesExist(files, token, status);

  std::vector<char*> translated_names;
  translated_names.reserve(files.size());
  for (int i = 0; i < files.size(); i++)

  outputs[0]->Unref();
}

TEST_P(CppGradients, TestSparseSoftmaxCrossEntropyWithLogitsGrad) {
  if (UseFunction()) {
    // TODO(b/168850692): Enable this.
    GTEST_SKIP() << "Can't take gradient of "
                    "SparseSoftmaxCrossEntropyWithLogits in tracing mode.";
  }

  // Score
  float X_vals[] = {1.0f, 2.0f, 3.0f, -5.0f, -4.0f, -3.0f, 2.0f, 0.0f, -1.0f};
  int64_t X_dims[] = {3, 3};

  // example they might require the root directory path to be in a special
  // format, etc. When we get there, we might decide to move this class to
  // `modular_filesystem_test.h` and extend the instantiation to also take as
  // argument an implementation for this method/a subclass factory (see
  // https://github.com/google/googletest/blob/master/googletest/docs/advanced.md#creating-value-parameterized-abstract-tests)