%arg2: !tf_res_md_f32 {tf.device = "/device:TPU:0"},
             %arg3: !tf_res_md_f32 {tf.device = "/device:TPU:1"}) {

    %0 = "tf.Const"() {value = dense<100> : tensor<i32>} : () -> tensor<i32>
    // CHECK: %[[STATE0:.*]] = "tf.VarHandleOp"()
    // CHECK-SAME: device = "/device:TPU:0"
    // CHECK: %[[STATE1:.*]] = "tf.VarHandleOp"()
    // CHECK-SAME: device = "/device:TPU:1"

  %3 = tfrt_fallback_async.executeop key(1) cost(512) device("/job:localhost/replica:0/task:0/device:CPU:0") "tf.AddV2"(%0, %0) {T = f32} : 1
  %4 = tfrt_fallback_async.executeop key(1) cost(512) device("/job:localhost/replica:0/task:0/device:CPU:0") "tf.AddV2"(%0, %0) {T = f32} : 1

      task_and_device = {task, device};
    }
  }

  return topology;
}

// Determine execution devices when topology and device assignment are defined.
// With a topology device coordinate to task and device mapping, device
// assignment device coordinates can then be mapped to task and device for TPU
// devices. The device assignment array is also validated.
//

    %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>

        // We assign default device to ops with device attribute that is empty.
        if (device_attr.getValue().empty()) {
          op->setAttr(kDeviceAttr, builder.getStringAttr(default_device_));
        }
      } else if (op->getDialect() == tf) {
        // Assign default device to all ops in Tensorflow dialect that do not
        // have device attribute.

  %20 = "tf.ExpandDims"(%19, %7) {device = ""} : (tensor<1xi32>, tensor<i32>) -> tensor<1x1xi32>
  %21 = "tf.Cast"(%20) {Truncate = false, device = ""} : (tensor<1x1xi32>) -> tensor<1x1xi64>
  %22 = "tf.Reshape"(%21, %12) {device = ""} : (tensor<1x1xi64>, tensor<1xi64>) -> tensor<1xi64>

// CHECK-SAME: ({{%.+}}: tensor<i64> {tf.device = "/job:localhost/replica:0/task:0/device:TPU:0"}, {{%.+}}: tensor<i32> {tf.device = "/job:localhost/replica:0/task:0/device:TPU:0"})
// CHECK-SAME: -> (tensor<i64> {tf.device = "/job:localhost/replica:0/task:0/device:TPU:0"}, tensor<i32> {tf.device = "/job:localhost/replica:0/task:0/device:TPU:0"})

  // CHECK: tac.cost = 1.000000e+03
  %0 = "tfl.add"(%arg0, %arg1) {fused_activation_function = "RELU", tac.device = "CPU"} : (tensor<10x10x10xf32>, tensor<10xf32>) -> tensor<10x10x10xf32>
  // CHECK: tac.cost = 1.000000e+03
  %1 = "tfl.mul"(%0, %arg1) {fused_activation_function = "RELU", tac.device = "CPU"} : (tensor<10x10x10xf32>, tensor<10xf32>) -> tensor<10x10x10xf32>
  func.return %1 : tensor<10x10x10xf32>
}

  if (!device_attr || device_attr.getValue().empty()) return true;

  DeviceNameUtils::ParsedName device;
  if (!DeviceNameUtils::ParseFullName(device_attr.getValue().str(), &device))
    return false;

  // We can't use GPU if operation explicitly placed on non-GPU device.
  return !device.has_type || device.type == ::tensorflow::DEVICE_GPU;
}

}  // namespace TF

  %cst = "tf.Const"() {device = "", value = dense_resource<__elided__> : tensor<2x3x3x2xbf16>} : () -> tensor<2x3x3x2xbf16>
  %cst_0 = "tf.Const"() {device = "", value = dense<[0, 3, 1, 2]> : tensor<4xi32>} : () -> tensor<4xi32>
  %cst_1 = "tf.Const"() {_tpu_replicate = "cluster", device = "", value = dense<[0, 2, 3, 1]> : tensor<4xi32>} : () -> tensor<4xi32>