int64_t bytes_transferred;
        bytes_transferred = block_fetcher_(key.first, key.second, block_size_,
                                           block->data.data(), status);
        block->mu.Lock();  // Reacquire the lock immediately afterwards
        if (TF_GetCode(status) == TF_OK) {
          block->data.resize(bytes_transferred, 0);
          // Shrink the data capacity to the actual size used.

  // otherwise be the name of this custom device. Ops are placed onto a custom
  // device if any of their inputs are on that custom device, but custom devices
  // are free to set a bad status in order to require explicit placement.
  void (*execute)(const TFE_Op* op, int* num_outputs,
                  TFE_TensorHandle** outputs, TF_Status* s, void* device_info);

  // Method to delete a device.

*   Include unit tests when you contribute new features, as they help to a)
    prove that your code works correctly, and b) guard against future breaking
    changes to lower the maintenance cost.
*   Bug fixes also generally require unit tests, because the presence of bugs
    usually indicates insufficient test coverage.
*   Keep API compatibility in mind when you change code in core TensorFlow,
    e.g., code in

      const auto func_op = GetFunc(ctx, default_value.func(), status);
      if (!status->status.ok()) return;
      // TODO(nareshmodi): TFE_OpSetAttrFunction and TFE_OpSetAttrFunctionList
      // require TFE_Op* and just convert it internally a NameAttrValue, so
      // consider adding an overload to the C API to make this case easier.
      TFE_OpSetAttrFunction(op, attr_name, func_op);
      TFE_DeleteOp(func_op);

// this to be called.
TF_CAPI_EXPORT void TF_InitMain(const char* usage, int* argc, char*** argv);

// Platform-specific implementation to return an unused port. (This should used
// in tests only.)
TF_CAPI_EXPORT int TF_PickUnusedPortOrDie(void);

// Fast path method that makes constructing a single scalar tensor require less
// overhead and copies.

  // Accumulate will forward op executions to the tape while the backward
  // function is running; this effectively adds the backward tape to the active
  // set (but does not require complicated callbacks to the language bindings).
  Status ForwardpropFromTape(
      const string& op_type, const std::vector<TapeTensor>& output_tensors,
      const std::function<BackwardFunction*()>& backward_function_getter,

  // parallel.
  //
  // Conceptually this is a thread pool with one thread per device. It requires
  // less synchronization than a thread pool would for this task, since Execute
  // acquires each thread in order (and so only one Execute will schedule
  // blocking collective operations at a time), and avoids some dynamic
  // allocation/scheduling.
  //

        body_nodes->push_back(node);
      } else {
        // This node is referenced in inputs. Currently, we place an
        // artificial restriction and require that when num_opers=-1, such
        // nodes must have a single output.
        if (node->num_outputs() != 1) {
          return InvalidArgument(
              "When `num_opers` is set to -1, nodes referenced in `inputs` "