%0 = "tf.Const"() {device = "/device:CPU:0", value = dense<0> : tensor<i64>} : () -> tensor<i64>
    %1 = "tf.Const"() {device = "/device:CPU:0", value = dense<5> : tensor<i64>} : () -> tensor<i64>
    %2 = "tf.Const"() {device = "/device:CPU:0", value = dense<1> : tensor<i64>} : () -> tensor<i64>

  return
}
```
### `-tf-device-assignment-by-func-attr`

_Device assignment in TF dialect using the device specified in the function attribute._

### `-tf-device-cluster-formation`

_Form clusters from instructions assigned to same device_

Clusters operations with the same device assignment id. For each

namespace mlir {
namespace TF {

void RuntimeDevices::AddDevice(const ParsedName& device) {
  device_names_.push_back(device);
}

void RuntimeDevices::AddGpuDevice(const ParsedName& device,
                                  const GpuDeviceMetadata& metadata) {
  device_names_.push_back(device);
  gpu_metadata_.insert({DeviceNameUtils::ParsedNameToString(device), metadata});
}

// CHECK: [[const:%.*]] = "tf.Const"() <{value = dense<0> : tensor<i32>}> {device = "/CPU:0"} : () -> tensor<i32>
// CHECK: [[x:%.*]] = "tf.AddV2"([[const]], [[const]]) {device = "/CPU:0"} : (tensor<i32>, tensor<i32>) -> tensor<i32>
// CHECK: "tf._TfrtSetResource"([[x]]) <{index = 0 : i64}> {device = "/CPU:0"} : (tensor<i32>) -> ()
// CHECK: [[const_1:%.*]] = "tf.Const"() <{value = dense<1> : tensor<i32>}> {device = "/CPU:0"} : () -> tensor<i32>

                            int num_underlying_devices,
                            TFE_CustomDevice* device, void** device_info) {
  device->copy_tensor_to_device = &CopyToParallelDevice;
  device->copy_tensor_from_device = &CopyTensorFromParallelDevice;
  device->delete_device = &DeleteParallelDevice;
  device->execute = &ParallelDeviceExecute;
  std::vector<std::string> underlying_devices_vector;

// CHECK: return %[[identity]] : tensor<1x3x2x2xf32>

func.func @cast_bf16_matmul_to_fp32(%arg0: tensor<1x10xf32>) -> (tensor<1x2xf32>) {
  %cst = "tf.Const"() {device = "", value = dense<1.000000e+01> : tensor<10x2xbf16>} : () -> tensor<10x2xbf16>
  %0 = "tf.Cast"(%arg0) {Truncate = false, device = ""} : (tensor<1x10xf32>) -> tensor<1x10xbf16>

  %value0 = "tf.LookupTableFindV2"(%handle, %key0, %default) {device = "/job:localhost/replica:0/task:0/device:CPU:0"} : (tensor<!tf_type.resource>, tensor<!tf_type.string>, tensor<i64>) -> tensor<i64>
  %cond = "tf.Equal"(%value0, %default) {device = "/job:localhost/replica:0/task:0/device:CPU:0", incompatible_shape_error = true} : (tensor<i64>, tensor<i64>) -> tensor<i1>

  tf_device.replicate() {n = 2 : i32, devices = {TPU_REPLICATED_CORE_0 = ["/DEVICE:0", "/DEVICE:1"], TPU_REPLICATED_CORE_1 = ["/DEVICE:2", "/DEVICE:3"]}} {
    tf_device.return
  }
  func.return

// CHECK:      tf_device.replicate
// CHECK-SAME: devices = {TPU_REPLICATED_CORE_0 = ["/DEVICE:0", "/DEVICE:1"], TPU_REPLICATED_CORE_1 = ["/DEVICE:2", "/DEVICE:3"]}
// CHECK-SAME: n = 2
// CHECK-NEXT:   tf_device.return
}

// `tf_device.launch`.
func.func @conflicting_device() {
  tf_executor.graph {
    %0 = tf_executor.island {
      // expected-error@+1 {{'tf_device.launch' op inner op has conflicting 'device' attribute, got 'GPU:0' but expected 'CPU:0'}}
      "tf_device.launch"() ({
        "tf.opA"() {device = "GPU:0"} : () -> ()
        tf_device.return
      }) {device = "CPU:0"} : () -> ()

inline constexpr absl::string_view kDeviceAssignmentAttr = "device_assignment";

// A TPU device for execution alongside its associated host CPU device.
struct TPUDeviceAndHost {
  TPUDeviceAndHost() = default;
  TPUDeviceAndHost(llvm::StringRef device, llvm::StringRef host)
      : device(device), host(host) {}

  std::string device;
  std::string host;
};