unsigned char TF_GetXlaConstantFoldingDisabled() {
  return static_cast<unsigned char>(
      tensorflow::GetBuildXlaOpsPassFlags()->tf_xla_disable_constant_folding);
}

void TF_SetXlaConstantFoldingDisabled(unsigned char should_enable) {
  tensorflow::GetBuildXlaOpsPassFlags()->tf_xla_disable_constant_folding =
      static_cast<bool>(should_enable);
}

void TF_SetXlaMinClusterSize(int size) {
  tensorflow::MarkForCompilationPassFlags* flags =

                                        std::move(mixed_handles), status.get());
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK);
  // Can't take the shape of a parallel tensor with varying numbers of axes, but
  // running operations on them is OK.
  TF_ASSERT_OK(unknown_length_vector->Shape(&shape));
  EXPECT_THAT(*shape, ElementsAre(-1));
  std::unique_ptr<TFE_Op, decltype(&TFE_DeleteOp)> size_op(

  ~DeviceThread();

  // Requests that the worker thread execute the specified operation. Blocks
  // until the previously pending operation (a StartExecute without a Join) has
  // finished, if any.
  //
  // `cancellation_manager` must live until after `Join` finishes and pending
  // `is_async` operations finish. In addition to allowing the caller to cancel
  // the operation, its `StartCancel` method will be called if op execution

                         std::vector<std::string>* control_ret_node_names,
                         bool* control_rets_updated) override {
    // Inject failure to function instantiation if finding a node that contains
    // the given node name (error_node_) and requested device (error_device_).
    for (const auto node : graph->get()->nodes()) {
      if (node->name().find(error_node_) != string::npos &&

  return nullptr;
}

// For TFE_CustomDevice::copy_tensor_from_device in the parallel device
// registration.
//
// Currently this is an error, and un-packing ParallelTensors must be performed
// explicitly by running a TPUReplicatedOutput operation on the parallel device.
//
// TODO(allenl): There are some use-cases that are only supported by copying to
// host at the moment (e.g. debug print on a tensor, .numpy(), etc.). We either

                                    /*remote_func_outputs=*/true);
}
TEST(CAPI, RemoteExecuteSilentCopiesLocalAsyncFuncOrdering) {
  // A remote input may be not ready when we start running a function. Test that
  // the function execution should wait until the remote input is ready.
  TestRemoteExecuteSilentCopiesFunc(/*async=*/true, /*remote=*/false,

  CONFIG_TOGGLE(function_optimization);
  CONFIG_TOGGLE(common_subgraph_elimination);
  CONFIG_TOGGLE(arithmetic_optimization);
  CONFIG_TOGGLE(debug_stripper);
  CONFIG_TOGGLE(constant_folding);
  CONFIG_TOGGLE(shape_optimization);
  CONFIG_TOGGLE(auto_mixed_precision);
  CONFIG_TOGGLE(auto_mixed_precision_onednn_bfloat16);
  CONFIG_TOGGLE(auto_mixed_precision_mkl);
  CONFIG_TOGGLE(pin_to_host_optimization);