// A sequence of thread wrappers, one per device, for executing operations in
  // 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

// tensors in `retvals`.
//
// 'retvals' must point to a pre-allocated array of TFE_TensorHandle* and
// '*num_retvals' should be set to the size of this array. It is an error if
// the size of 'retvals' is less than the number of outputs. This call sets
// *num_retvals to the number of outputs.
//
// If async execution is enabled, the call may simply enqueue the execution

// 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.
TF_CAPI_EXPORT extern TFE_TensorHandle* TFE_NewTensorHandleFromScalar(
    TF_DataType data_type, void* data, size_t len, TF_Status* status);

// Specify the server_def that enables collective ops.

  // there isn't an intervening PushState. This is useful for ordering
  // ForwardAccumulators, where more deeply nested accumulators should not see
  // computations from less deeply nested accumulators.
  bool BusyAccumulating() const { return call_state_.top().accumulating; }

  // Fetches the current Jacobian-vector product associated with `tensor_id`, or
  // a nullptr if none is available.