limitations under the License.
==============================================================================*/

// MakeCleanup(f) returns an RAII cleanup object that calls 'f' in its
// destructor. The easiest way to use MakeCleanup is with a lambda argument,
// capturing the return value in an 'auto' local variable. Most users will not
// need more sophisticated syntax than that.
//
// Example:

  TFE_Execute(op.get(), &var_handle, &num_retvals, status.get());
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  ASSERT_TRUE(executed);
  auto handle_cleaner = tensorflow::gtl::MakeCleanup(
      [var_handle]() { TFE_DeleteTensorHandle(var_handle); });

  // Assign to the variable, copying to the custom device.
  std::unique_ptr<TFE_TensorHandle, decltype(&TFE_DeleteTensorHandle)> one(

  Status s = InitialGradients(vspace, target_tensor_ids,
                              sources_that_are_targets, output_gradients,
                              tensor_tape_, state.op_tape, &gradients);
  auto cleanup = gtl::MakeCleanup([this, &state]() {
    if (!persistent_) {
      // Release all backprop functions
      for (const auto& pair : state.op_tape) {
        pair.second.backward_function_deleter(pair.second.backward_function);
      }

                                   const std::shared_ptr<Block>& block,
                                   TF_Status* status) {
  bool downloaded_block = false;
  auto reconcile_state = MakeCleanup([this, &downloaded_block, &key, &block] {
    // Perform this action in a cleanup callback to avoid locking mu_ after
    // locking block->mu.
    if (downloaded_block) {
      absl::MutexLock l(&mu_);