TF_Graph* const cond_graph;
  const TF_Output* const cond_inputs;
  TF_Output cond_output;

  // The loop body graph. The inputs are the current values of the loop
  // variables. The outputs are the updated values of the loop variables.
  TF_Graph* const body_graph;
  const TF_Output* const body_inputs;
  TF_Output* const body_outputs;

  TF_SessionOptions* opts = TF_NewSessionOptions();
  TF_Session* sess = TF_NewSession(graph, opts, s);
  TF_DeleteSessionOptions(opts);

  TF_Output feeds[] = {TF_Output{a, 0}, TF_Output{b, 0}};
  TF_Output fetches[] = {TF_Output{plus2, 0}, TF_Output{plusB, 0}};

  const char* handle = nullptr;
  TF_SessionPRunSetup(sess, feeds, TF_ARRAYSIZE(feeds), fetches,

    for (int i = 0; i < expected.size(); ++i) {
      ASSERT_EQ(expected[i], data[i]) << "Different data at index " << i;
    }
  }

  std::vector<TF_Output> ToOutput(const std::vector<TF_Operation*> ops) {
    std::vector<TF_Output> out;
    for (auto op : ops) {
      out.push_back({op, 0});
    }
    return out;
  }

  void Define(int num_opers, const std::vector<TF_Operation*>& opers,

  // CHECK-NEXT:     "tf.Yield"(%[[C_OUTPUT]], %[[D_OUTPUT]])
  // CHECK-NOT:    "tf.IfRegion"
  // CHECK         "tf.AssignVariableOp(arg0, %[[IF_OUTPUT#0]])
  // CHECK         "tf.E"(%[[IF_OUTPUT#1]])
  // CHECK-NEXT    "tf.F"(%[[IF_OUTPUT#1]])
  // CHECK-NOT:    "tf.IfRegion"
  "tf_device.cluster"() ({
    %0 = "tf.Const"() {value = dense<true> : tensor<i1>} : () -> tensor<i1>
    %1 = "tf.IfRegion"(%0) ({

    Res<Variadic<TF_Tensor>, [{A list of tensors that will be returned to the device.}]>:$outputs
  );

  TF_DerivedOperandTypeListAttr Tinputs = TF_DerivedOperandTypeListAttr<0>;
  TF_DerivedResultTypeListAttr Toutputs = TF_DerivedResultTypeListAttr<0>;
}

def TF_ConfigureAndInitializeGlobalTPUOp : TF_Op<"ConfigureAndInitializeGlobalTPU", []> {
  let summary = [{
An op that initialize the TPU system in a multi-client set up.