// CHECK: %[[LAUNCH_OUTPUT:[0-9]*]] = "tf_device.launch_func"(%[[A_OUTPUT]]) <{device = "/device:test_device:0", func = @[[LAUNCH:.*]]}>
      %3 = "tf_device.launch"() ({
        %4 = "tf.B"(%2) : (tensor<?xi32>) -> tensor<?xi32>
        tf_device.return %4 : tensor<?xi32>
      }) {device = "/device:test_device:0"} : () -> tensor<?xi32>

      // CHECK: tf_executor.yield %[[LAUNCH_OUTPUT]]
      tf_executor.yield %3 : tensor<?xi32>
    }

namespace mlir {
namespace TF {

namespace {

const char kDeviceAttr[] = "device";
const char kDeviceGpu[] = "GPU";

std::optional<std::string> GetOpDevice(mlir::Operation *op) {
  mlir::StringAttr device = op->getAttrOfType<mlir::StringAttr>(kDeviceAttr);
  if (!device || device.getValue().empty()) {
    return std::nullopt;
  }
  tensorflow::DeviceNameUtils::ParsedName parsed_name;

        "Cannot get device metadata from non-PJRT device \"", device->name(),
        "\". GetMetadata must only be called on a device derived from "
        "PjRtBaseDevice. Either an internal bug has been triggered, or an "
        "XLA-specific op has been placed on the wrong device.");
  }
  return &pjrt_device->metadata_;
}

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);

    Op<TFRT_GPU_Dialect, mnemonic, traits> {
}

// TODO(b/260267885): We may add a device argument when we want to support
// GPU MIG.
def TransferToDeviceOp: Gpu_Op<"transfer_to_device"> {
  let summary = "Transfer a CPU tensor to device.";

  let description = [{
    Transfer a CPU tensor to device.

    Example:
      %device_tensor = gpurt.transfer_to_device %cpu_tensor
  }];

  };

  devices.ForEach([&](jit::DeviceId device) {
    if (device_info_cache.IsGpu(device)) {
      if (maybe_gpu_device) {
        multiple_gpu_devices = is_multiple_devices(device, &maybe_gpu_device);
        if (multiple_gpu_devices) return false;
      } else {
        maybe_gpu_device = device;
      }
    } else if (device_info_cache.IsCpu(device)) {
      if (maybe_cpu_device) {

    %1 = "tf.Rank"(%0) {_tpu_replicate = "cluster", device = ""} : (tensor<*xi1>) -> tensor<*xi32>
    %2 = "tf.Range"(%cst_0, %1, %cst_1) {_tpu_replicate = "cluster", _xla_outside_compilation = "0", device = ""} : (tensor<i32>, tensor<*xi32>, tensor<i32>) -> tensor<*xi32>
    %3 = "tf.All"(%0, %2) {_tpu_replicate = "cluster", _xla_outside_compilation = "0", device = "", keep_dims = false} : (tensor<*xi1>, tensor<*xi32>) -> tensor<*xi1>

     %3 = builtin.unrealized_conversion_cast to tensor<i32>
     // CHECK: TPUCopyWithDynamicShape{{.*}}device = "foobar"
     %4, %5 = "tf.TPUCopyWithDynamicShape"(%0, %1, %2, %3) {operandSegmentSizes = array<i32: 2, 2>} : (tensor<2048xi32>, tensor<2048xi32>, tensor<i32>, tensor<i32>) -> (tensor<2048xi32>, tensor<2048xi32>)
    "tf.TPUExecute"(%4, %arg0) {device = "foobar"} : (tensor<2048xi32>, tensor<!tf_type.string>) -> ()
    return
}

// -----

      // CHECK-SAME: device = "/device:test_device:0"
      %3 = "tf_device.launch_func"(%2) {device = "/device:test_device:0", func = @_func} : (tensor<?xf32>) -> tensor<?xf32>

      // CHECK: %[[CALL_OUTPUT_1:[0-9]*]] = "tf.PartitionedCall"(%[[CALL_OUTPUT_0]])
      // CHECK-SAME: f = @_func
      // CHECK-SAME: device = "/device:test_device:1"

def TFXLADeviceSpecificTransforms : Pass<"tfxla-device-specific-transforms",
                                            "mlir::func::FuncOp"> {
  let summary = "Transforms ops that require device context into device independent TF Ops.";

  let description = [{"Transforms device specific ops into device independent"
                    "ops."}];

  let options = [
    Option<"device_type_", "device-type", "std::string",