// 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);
    for (int i = 0; i < num_inputs; ++i) {

    int64_t target_id = target_tensor_ids[i];
    if (sources_set.find(target_id) != sources_set.end()) {
      auto tensor = targets[i];
      sources_that_are_targets.insert(
          std::make_pair(target_id, TapeTensor(tensor)));
    }
  }

  TF_RETURN_IF_ERROR(GradientTape::ComputeGradient(
      vspace, target_tensor_ids, source_tensor_ids, sources_that_are_targets,

  TFE_TensorHandle* h0 = TestVariable(ctx, 1.0, task1_name);
  TFE_TensorHandle* h1 = TestVariable(ctx, 2.0, task2_name);
  TFE_TensorHandle* h2 = TestVariable(ctx, 3.0, task0_name);

  // Add a sync point to make sure that variables have been initialized
  // before the function execution starts.
  TFE_ContextAsyncWait(ctx, status);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

ParallelDevice::ParallelDevice(const std::vector<std::string>& devices,
                               bool is_async, int in_flight_nodes_limit)
    : underlying_devices_(devices),
      default_cancellation_manager_(absl::make_unique<CancellationManager>()) {
  device_threads_.reserve(devices.size());
  for (int device_index = 0; device_index < devices.size(); ++device_index) {
    device_threads_.emplace_back(new DeviceThread(

}

// Converts `ninputs` and `inputs` into `inputs_tensors` and `input_nodes` and
// does various checks while doing so. `input_nodes` will contain the same
// information as input_tensors just in a different structure to make
// following processing easier. TODO(iga): Simplify this nested structure.
Status ProcessInputs(
    const TF_Graph* fn_body, const char* fn_name, int ninputs,
    const TF_Output* inputs, std::vector<OutputTensor>* input_tensors,

      /*run_eager_op_as_function=*/opts->run_eager_op_as_function,
      /*jit_compile_rewrite=*/opts->jit_compile_rewrite);
#if !defined(IS_MOBILE_PLATFORM)
  eager_context->SetDistributedManager(
      std::make_unique<tensorflow::EagerContextDistributedManager>(
          eager_context));
#endif  // !IS_MOBILE_PLATFORM
  return tensorflow::wrap(eager_context);
}

void TFE_DeleteContext(TFE_Context* ctx) {

  file_block_cache = std::make_unique<RamFileBlockCache>(
      block_size, max_bytes, max_staleness,
      [this](const std::string& filename, size_t offset, size_t buffer_size,
             char* buffer, TF_Status* status) {
        return LoadBufferFromGCS(filename, offset, buffer_size, buffer, this,
                                 status);
      });
  stat_cache = std::make_unique<ExpiringLRUCache<GcsFileStat>>(

  auto file = std::make_unique<TF_RandomAccessFile>();
  std::string translated_name = TranslateName(fname);
  ops_->new_random_access_file(filesystem_.get(), translated_name.c_str(),
                               file.get(), plugin_status.get());

  if (TF_GetCode(plugin_status.get()) == TF_OK)
    *result = std::make_unique<ModularRandomAccessFile>(