// CHECK-NOT:        "tf._TPUDeviceOrdinalPlaceholder"
    // CHECK:            %[[B_OUTPUT:[0-9]*]] = "tf.B"()
    // CHECK:            "tf._XlaSendFromHost"(%[[B_OUTPUT]], %[[PROGRAM_OUTPUT]])
    // CHECK-SAME:       device_ordinal = 0
    // CHECK-SAME:       device_type = "TPU"
    // CHECK-SAME:       key = "host_compute_channel_0_retvals"

  for (Node* n : g->nodes()) {
    if (n->type_string() == "_XlaSendFromHost") {
      send_from_host = n;
      break;
    }
  }
  if (!send_from_host) {
    return errors::Internal("Shape inference graph ",
                            shape_inference_graph_name,
                            " does not have _XlaSendFromHost node.");
  }

  // See if the SendFromHost node exists in `host_graph`.

  string name = absl::StrCat("outside_compilation_", cluster, "_",
                             new_func_name, "_", oc_cluster, "_send");
  NodeBuilder node_builder(opts.WithName(name).GetNameForOp("_XlaSendFromHost"),
                           "_XlaSendFromHost", opts.op_registry());
  node_builder.Input(inputs);
  node_builder.Input(std::move(key_input));
  std::vector<DataType> dtypes;
  for (const auto& node : inputs) {

    // CHECK: %5 = "tf.Shape"(%4) {_xla_outside_compilation = "0"} : (tensor<1120x8xi32>) -> tensor<2xi32>
    // CHECK: "tf._XlaSendFromHost"(%3#1, %3#2, %4, %5, %0) <{device_ordinal = 0 : i64, device_type = "TPU", key = "host_compute_channel_0_retvals"}> : (tensor<1x1120x8xi32>, tensor<1x1120x8xi32>, tensor<1120x8xi32>, tensor<2xi32>, tensor<3x!tf_type.string>) -> ()