xla::ExecutableBuildOptions build_options;
  if (result.collective_info) {
    build_options.set_num_replicas(result.collective_info->group_size);
  }
  build_options.set_device_ordinal(
      options.device_ordinal != -1 ? options.device_ordinal
                                   : local_client->default_device_ordinal());
  build_options.set_result_layout(result.xla_output_shape);
  build_options.set_device_allocator(options.device_allocator.get());

    xla::LocalClient* client, se::DeviceMemoryAllocator* allocator) {
  se::Stream* stream = GetStream(ctx);
  int device_ordinal = stream ? stream->parent()->device_ordinal()
                              : client->default_device_ordinal();
  XlaComputationLaunchContext launch_context(
      client, allocator, device_ordinal,
      /*allocate_xla_tensors=*/platform_info.is_on_xla_device(),

// CHECK:      tf_executor.fetch


// Tests tf._TPUDeviceOrdinalPlaceholder ops are replaced with explicit device
// ordinal constant values given by logical_core.
// CHECK-LABEL: func @device_ordinals
func.func @device_ordinals() {
  tf_executor.graph {
    %0:1 = tf_executor.island {

          %2 = "tf._XlaRecvAtHost"(%1) {device_ordinal = 0 : i64, key = "host_compute_channel_0_0_args"} : (tensor<3x!tf_type.string>) -> tensor<f32>
          %3 = "tf.Identity"(%2) : (tensor<f32>) -> tensor<f32>
          "tf._XlaSendFromHost"(%3, %1) {device_ordinal = 0 : i64, key = "host_compute_channel_0_0_retvals"} : (tensor<f32>, tensor<3x!tf_type.string>) -> ()

  Output x = ops::Placeholder(root.WithOpName("x"), DT_FLOAT);
  NameAttrList f_name_attr;
  f_name_attr.set_name("XTimesTwoFloat");
  ops::TPUPartitionedCall f(root.WithOpName("f"), {x}, /*device_ordinal=*/0,
                            {DT_FLOAT}, f_name_attr);

  TF_ASSERT_OK(root.ToGraph(&graph));
  EXPECT_TRUE(IsInferenceGraph(graph, /*function_library=*/nullptr));
}

        "tf.EnqueueTPUEmbeddingRaggedTensorBatch"(%arg0) {table_ids = [1, 2], device_ordinal = 1} : (tensor<!tf_type.string>) -> ()
        // expected-remark@above {{ID: 0}}
        // expected-remark@above {{Successors: {1}}}
        "tf.EnqueueTPUEmbeddingRaggedTensorBatch"(%arg0) {table_ids = [1, 2], device_ordinal = 1} : (tensor<!tf_type.string>) -> ()
        // expected-remark@above {{ID: 1}}

    ConfinedAttr<TypeArrayAttr, [ArrayMinCount<1>]>:$dtypes,
    TF_ShapeAttrArray:$shapes,
    DefaultValuedOptionalAttr<I64ArrayAttr, "{}">:$layouts,
    DefaultValuedOptionalAttr<I64Attr, "-1">:$device_ordinal
  );

  let results = (outs);
}

// This op is manually defined because the attribute name `template` (which is
// a keyword) is changed to `strtemplate`.

    // CHECK: %0 = "tf._XlaCompileMlirPlaceholderProgramKey"() : () -> tensor<3x!tf_type.string>
    // CHECK: %1:4 = "tf._XlaRecvAtHost"(%0) <{device_ordinal = 0 : i64, device_type = "TPU", key = "host_compute_channel_0_args"}> : (tensor<3x!tf_type.string>) -> (tensor<3360x8xi32>, tensor<3xi32>, tensor<i32>, tensor<2xi32>)

                           "_XlaRecvAtHost", opts.op_registry());
  node_builder.Input(std::move(key_input));
  return opts.WithAttr("Toutputs", dtypes)
      .WithAttr("key", key)
      .WithAttr("device_ordinal", 0)
      .WithAttr("_encapsulate", cluster)
      .WithAttr("_outside", oc_cluster)
      .FinalizeBuilder(&node_builder);
}

Node* SendFromHost(ops::NodeOut key_input, const string& cluster,

// Helper function to get an absolute device string, combining device and
// ordinal attribute values.
std::string GetAbsDeviceStr(Operation* op, uint64_t device_ordinal) {
  std::string device_ordinal_str = std::to_string(device_ordinal);
  auto device_attr = op->getAttrOfType<StringAttr>("device");
  if (!device_attr || device_attr.getValue().empty()) return device_ordinal_str;