static const char* const module_str =
      R"(
module attributes {tf.devices = ["/job:worker/replica:0/task:0/device:CPU:0", "/job:worker/replica:0/task:0/device:GPU:0"]} {
  func.func @convert_cluster_func(%arg0: tensor<i32>) -> () {
    %2 = "tf_device.parallel_execute"() ({

      %3 = "tf_device.cluster_func"(%arg0) {device = "/job:localhost/replica:0/task:0/device:GPU:0", func = @func} : (tensor<i32>) -> tensor<i32>

            "Switch a device into 'on-demand' mode, where instead of "
            "autoclustering ops are compiled one by one just-in-time."),

       Flag("tf_xla_enable_xla_devices",
            &device_flags->tf_xla_enable_xla_devices,
            "Generate XLA_* devices, where placing a computation on such a "
            "device"
            "forces compilation by XLA. Deprecated."),

  }
}

ParallelDevice::ParallelDevice(const std::vector<std::string>& devices,
                               bool is_async, int in_flight_nodes_limit)
    : underlying_devices_(devices),
      default_cancellation_manager_(absl::make_unique<CancellationManager>()) {
  device_threads_.reserve(devices.size());
  for (int device_index = 0; device_index < devices.size(); ++device_index) {
    device_threads_.emplace_back(new DeviceThread(

                                  std::unique_ptr<Graph>* result) {
  std::vector<std::unique_ptr<Device>> devices;
  devices.push_back(FakeDevice::Make(kDeviceName, DEVICE_GPU));
  devices.push_back(FakeDevice::Make(kHostName, DEVICE_CPU));

  std::unique_ptr<DeviceSet> device_set(new DeviceSet());
  for (auto& device : devices) {
    device_set->AddDevice(device.get());
  }

  auto graph = std::make_unique<Graph>(OpRegistry::Global());

    const Cluster& cluster_a, const Cluster& cluster_b) {
  DeviceSet devices = cluster_a.devices();
  devices.UnionWith(cluster_b.devices());

  TF_ASSIGN_OR_RETURN(
      std::optional<jit::DeviceId> maybe_chosen_device,
      MaybePickDeviceForXla(device_info_cache_, devices,
                            /*allow_mixing_unknown_and_cpu=*/false));
  if (!maybe_chosen_device.has_value()) {

LogicalResult GetTpuDeviceAssignment(
    ClusterOp cluster, ReplicateOp replicate, mlir::ModuleOp module,
    absl::StatusOr<TPUDeviceAssignment>& status_or_tpu_device_assignment) {
  mlir::TF::RuntimeDevices devices;
  if (failed(tensorflow::GetDevicesFromOp(module, &devices))) return failure();

  uint32_t num_replicas = replicate.getN();

  auto num_cores_per_replica_attr = cluster->getAttrOfType<mlir::IntegerAttr>(

// CHECK-NEXT: [[o_chain:%.*]], [[o:%.*]] = tfrt_fallback_async.executeop.seq([[in_chain]]) key(0) cost({{.*}}) device("/job:localhost/replica:0/task:0/device:CPU:0") "tf.VarHandleOp"()
// CHECK-NEXT: [[o_chain_0:%.*]], [[o1:%.*]] = tfrt_fallback_async.executeop.seq([[in_chain]]) key(1) cost({{.*}}) device("/job:localhost/replica:0/task:0/device:CPU:0") "tf.ReadVariableOp"([[o]]) {dtype = f32} : 1

)

// HostUtils defines the set of methods for interacting with paths on a host.
type HostUtils interface {
	// DeviceOpened determines if the device (e.g. /dev/sdc) is in use elsewhere
	// on the system, i.e. still mounted.
	DeviceOpened(pathname string) (bool, error)
	// PathIsDevice determines if a path is a device.
	PathIsDevice(pathname string) (bool, error)

// "replicate_on_last_tile_dim" and "last_tile_dims" should be deducted from the
// real Tensor dimensions when tiled.
// For example:
// f32[8,512](sharding={devices=[1,1,2]0,1 last_tile_dims={REPLICATED})
// also means a replicated tensor over all devices.
//
// See xla_data.proto for detailed explanations on the fields.
int GetDimsFromXLAShardingTiled(const xla::OpSharding& xla_sharding);

// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.

#include "textflag.h"

// Linux/ARM atomic operations.

// Because there is so much variation in ARM devices,
// the Linux kernel provides an appropriate compare-and-swap
// implementation at address 0xffff0fc0.  Caller sets:
//	R0 = old value
//	R1 = new value
//	R2 = addr
//	LR = return address