const char* first_device = "/job:worker/replica:0/task:1/device:CPU:0";
  const char* second_device = "/job:worker/replica:0/task:2/device:CPU:0";
  const char* device_name = "/job:localhost/replica:0/task:0/device:CUSTOM:0";
  std::array<const char*, 2> underlying_devices{first_device, second_device};
  RegisterParallelDevice(context.get(), device_name, underlying_devices,
                         status.get());

  auto iter = name_ranges.find(arg_name);
  if (iter == name_ranges.end()) {
    status->status = InvalidArgument("Output arg '", arg_name, "' not found");
    return -1;
  }
  return iter->second.second - iter->second.first;
}

int TF_OperationNumInputs(TF_Operation* oper) {
  return oper->node.num_inputs();
}

TF_DataType TF_OperationInputType(TF_Input oper_in) {

  GetFactories()[name] = factory;
}

Status SetDefaultTracingEngine(const char* name) {
  auto entry = GetFactories().find(name);
  if (entry != GetFactories().end()) {
    default_factory = GetFactories().find(name)->second;
    return absl::OkStatus();
  }
  string msg = absl::StrCat(
      "No tracing engine factory has been registered with the key '", name,
      "' (available: ");

  // Another renaming of local device
  const string second_name = "def";
  server_def.set_job_name(second_name);
  server_def.mutable_cluster()->mutable_job(0)->set_name(second_name);
  (*server_def.mutable_cluster()->mutable_job(0)->mutable_tasks())[0] =
      absl::StrCat(second_name, ":",
                   tensorflow::testing::PickUnusedPortOrDie());

  tensorflow::Env* env = tensorflow::Env::Default();
  {
    tensorflow::mutex_lock lock(mu);
    auto it = loaded_libs->find(library_filename);
    if (it != loaded_libs->end()) {
      lib_handle->lib_handle = it->second;
    } else {
      status->status =
          env->LoadDynamicLibrary(library_filename, &lib_handle->lib_handle);
      if (status->status.ok()) {
        TF_CHECK_OK(

  if (iter == registry_.end()) {
    const string error_msg = "No gradient defined for op: " + op.op_name + ".";
    return errors::NotFound(error_msg);
  }
  gradient_function->reset(iter->second(op));
  return absl::OkStatus();
}

TapeTensor::TapeTensor(AbstractTensorHandle* handle) : handle_(handle) {
  handle_->Ref();
}
TapeTensor::TapeTensor(const TapeTensor& other) {
  handle_ = other.handle_;

  block_size = kDefaultBlockSize;
  size_t max_bytes = kDefaultMaxCacheSize;
  uint64_t max_staleness = kDefaultMaxStaleness;

  // Apply the overrides for the block size (MB), max bytes (MB), and max
  // staleness (seconds) if provided.
  const char* block_size_env = std::getenv(kBlockSize);
  if (block_size_env && absl::SimpleAtoi(block_size_env, &value)) {
    block_size = value * 1024 * 1024;
  }

                .insert(EncodeTensorNameSlice(
                    string(v2_reader_->key()) /* full var's name */,
                    TensorSlice(slice_proto)))
                .second);
    }
  }

  // Second pass: adds the entries, ignoring the filtered keys.
  std::unique_ptr<TensorSliceReader::VarToShapeMap> var_to_shape_map(
      new TensorSliceReader::VarToShapeMap);

  tensorflow::AttrValueMap::const_iterator attr_found = attr_values.find("T");
  EXPECT_NE(attr_found, attr_values.cend());
  EXPECT_EQ(attr_found->second.list().type(0), tensorflow::DataType::DT_BOOL);
  EXPECT_EQ(attr_found->second.list().type(1), tensorflow::DataType::DT_FLOAT);
  EXPECT_EQ(attr_found->second.list().type(2), tensorflow::DataType::DT_INT32);

  TFE_TensorHandle* retvals[1] = {nullptr};
  int num_retvals = 1;

      "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;
  backprop->Unref();
  return absl::OkStatus();
}

Status BiasAddModel(AbstractContext* ctx,