mlir::Builder builder(ctx);

  // Collect devices with attached metadata.
  llvm::SmallVector<mlir::NamedAttribute, 8> devices;
  devices.reserve(device_set->devices().size());

  // For device that do not have any metadata, or if we failed to parse metadata
  // from the DeviceSet, we add a unit attribute to the `tf.devices` attribute.
  for (Device* device : device_set->devices()) {

  llvm::SmallVector<std::unique_ptr<Device>, 2> devices;
  devices.push_back(FakeDevice::Make(cpu0));
  devices.push_back(FakeDevice::Make(gpu0, "compute capability: 7.0"));
  devices.push_back(FakeDevice::Make(gpu1));

  DeviceSet device_set;
  for (auto& device : devices) device_set.AddDevice(device.get());
  AddDevicesToOp(*module_ref, &device_set);

  auto devices_attr =

func New(devEntries []PodDevicesEntry, devices map[string][]string) DeviceManagerCheckpoint {
	return newV2(devEntries, devices)
}

func newV2(devEntries []PodDevicesEntry, devices map[string][]string) DeviceManagerCheckpoint {
	return &Data{
		Data: checkpointData{
			PodDeviceEntries:  devEntries,
			RegisteredDevices: devices,
		},
	}
}

// MarshalCheckpoint returns marshalled data

};

// Caches some miscellaneous information about TF devices.  Thread compatible.
class DeviceInfoCache {
 public:
  bool IsGpu(DeviceId device) const { return is_gpu_[device.id()]; }
  bool IsCpu(DeviceId device) const { return is_cpu_[device.id()]; }

  absl::string_view GetNameFor(DeviceId device) const {
    return names_[device.id()];
  }

  absl::StatusOr<DeviceId> GetIdFor(absl::string_view name);

    SessionOptions options;
    std::vector<std::unique_ptr<Device>> devices;
    Status s = device_factory->CreateDevices(
        options, "/job:worker/replica:0/task:0", &devices);
    device_ = std::move(devices[0]);

    AllocatorAttributes host_alloc_attr;
    host_alloc_attr.set_on_host(true);
    host_allocator_ = device_->GetAllocator(host_alloc_attr);

    AllocatorAttributes device_alloc_attr;

func (info *ClaimInfo) cdiDevicesAsList() []kubecontainer.CDIDevice {
	var cdiDevices []kubecontainer.CDIDevice
	for _, devices := range info.CDIDevices {
		for _, device := range devices {
			cdiDevices = append(cdiDevices, kubecontainer.CDIDevice{Name: device})
		}
	}
	return cdiDevices
}

// addPodReference adds a pod reference to the claim info.

  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>

// 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

	PodUIDs sets.Set[string]

	// ResourceHandles is a list of opaque resource data for processing by a specific kubelet plugin
	ResourceHandles []resourcev1alpha2.ResourceHandle

	// CDIDevices is a map of DriverName --> CDI devices returned by the
	// GRPC API call NodePrepareResource
	CDIDevices map[string][]string
}

// ClaimInfoStateWithoutResourceHandles is an old implementation of the ClaimInfoState
// TODO: remove in Beta

// RUN: tf-opt %s -split-input-file -tf-xla-broadcast | FileCheck %s