// The name of the compilation device (e.g., "XLA_CPU_JIT");
    string compilation_device_name;

    // If 'use_multiple_streams' is true, we create separate streams for
    // compute, host-to-device, and device-to-host communication.
    bool use_multiple_streams = false;

    // If true, the XLA devices with the same device ordinal will share the same
    // compute stream. Otherwise each XLA device will having their own compute

"909", tf.device = "/job:tpu_host_worker/replica:0/task:0/device:CPU:0"}, %arg7: tensor<*x!tf_type.resource<tensor<25001x64xf32>>> {tf._user_specified_name = "911", tf.device = "/job:tpu_host_worker/replica:0/task:0/device:CPU:0"}, %arg8: tensor<*x!tf_type.resource<tensor<25001x64xf32>>> {tf._user_specified_name = "913", tf.device = "/job:tpu_host_worker/replica:0/task:1/device:CPU:0"}, %arg9: tensor<*x!tf_type.resource<tensor<25001x64xf32>>> {tf._user_specified_name = "915", tf.device = "/job:t...

    %0 = tfl.add %arg0, %arg1 {fused_activation_function = "RELU6", tac.device = "GPU", tac.inference_type = "FLOAT"} : tensor<1xf32>
    %1 = tfl.mul %0, %arg2 {fused_activation_function = "RELU6", tac.device = "GPU", tac.inference_type = "FLOAT"} : tensor<1xf32>
    func.return %1 : tensor<1xf32>

    %cst = "tf.Const"() <{value = dense<[0, 1]> : tensor<2xi32>}> {device = ""} : () -> tensor<2xi32>
    %cst_0 = "tf.Const"() <{value = dense<1.000000e+01> : tensor<f32>}> {device = ""} : () -> tensor<f32>
    %0 = "tf.Sum"(%arg0, %cst) <{keep_dims = false}> {device = ""} : (tensor<1x4xf32>, tensor<2xi32>) -> tensor<f32>
    %1 = "tf.Greater"(%0, %cst_0) {device = ""} : (tensor<f32>, tensor<f32>) -> tensor<i1>

// and ensures that an upto date state (e.g. number of devices and device health)
// is captured. Also, registered device and device to container allocation
// information is checkpointed to the disk.
func (m *ManagerImpl) PluginListAndWatchReceiver(resourceName string, resp *pluginapi.ListAndWatchResponse) {
	var devices []pluginapi.Device
	for _, d := range resp.Devices {

                                   TF_Status* status) {
    params->device->hardware_name = hardware_name;
    params->device->device_vendor = vendor;
    params->device->pci_bus_id = pci_bus_id;
  };

  device_fns_.get_numa_node = [](const SP_Device* device) { return 123; };
  device_fns_.get_memory_bandwidth = [](const SP_Device* device) -> int64_t {
    return 54;
  };

  %0 = "tf.Conv2D"(%arg0, %arg1) {
    data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [],
    padding = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true
  } : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<*xf32>
  %1 = "tf.BiasAdd"(%0, %cst) {data_format = "NHWC", device = ""} : (tensor<*xf32>, tensor<2xf32>) -> tensor<*xf32>
  %2 = "tf.Relu6"(%1) {device = ""} : (tensor<*xf32>) -> tensor<*xf32>

}

def LaunchToDeviceAttributePass : Pass<"tf-launch-to-device-attribute", "mlir::func::FuncOp"> {
  let summary = "Hoists and annotates device launch inner ops with associated device attribute.";

  let description = [{
    This pass hoists a `tf_device.launch` body and assigns a `device` attribute
    to each TensorFlow dialect op in the body based on the `device` attribute on

    Device* device = dynamic_cast<Device*>(ctx->device());
    if (device == nullptr) {
      return errors::Internal("DeviceBase was not a Device.");
    }
    ctx->op_device_context()->CopyCPUTensorToDevice(
        &const_tensor, device, output_tensor,
        [&](Status status) { TF_CHECK_OK(status); });

    if (device->device_type() == DEVICE_GPU) {

    %cst = "tf.Const"() <{value = dense<[0, 1]> : tensor<2xi32>}> {device = ""} : () -> tensor<2xi32>
    %cst_0 = "tf.Const"() <{value = dense<1.000000e+01> : tensor<f32>}> {device = ""} : () -> tensor<f32>
    %0 = "tf.Sum"(%arg0, %cst) <{keep_dims = false}> {device = ""} : (tensor<1x4xf32>, tensor<2xi32>) -> tensor<f32>
    %1 = "tf.Greater"(%0, %cst_0) {device = ""} : (tensor<f32>, tensor<f32>) -> tensor<i1>