// caller of event_mgr->ThenExecute() blocks on the completion of the callback
  // (as in the case of ConstOp kernel creation on GPU, which involves copying a
  // CPU tensor to GPU).
  // Setting a larger thread pool does not help with the Swift caller, as we use
  // a different TFE context for each thread of execution (for running graph
  // functions, and their send/recvs corountines).

// 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,

  if (TF_GetCode(status) != TF_OK) return nullptr;
  if (parallel_tensor->num_tensors() == 1) {
    // Copy-off for single-device tensors is allowed to make debugging dynamic
    // control flow easier.
    return TFE_TensorHandleCopySharingTensor(parallel_tensor->tensor(0),
                                             status);
  } else {
    TF_SetStatus(
        status, TF_UNIMPLEMENTED,
        absl::StrCat(

    }
    *state_iter = std::move(s);
    ++state_iter;
  }
  status->status = absl::OkStatus();
}

void TFE_WaitAtBarrier(TFE_Context* ctx, const char* barrier_id,
                       int64_t barrier_timeout_in_ms, TF_Status* status) {
  tensorflow::ImmediateExecutionDistributedManager* dist_mgr =
      tensorflow::unwrap(ctx)->GetDistributedManager();
  tsl::CoordinationServiceAgent* coord_agent =

      // 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);
    } break;
    case tensorflow::AttrValue::kList: {
      // String