%outputs, %control = tf_executor.island wraps "tf.GuaranteeConst"(%arg1) {T = f32, device = ""} : (tensor<f32>) -> tensor<f32>
    %outputs_0, %control_1 = tf_executor.island wraps "tf.GuaranteeConst"(%arg2) {T = f32, device = ""} : (tensor<f32>) -> tensor<f32>
    %control_2 = tf_executor.island wraps "tf.NoOp"() {_pivot_for_cluster = "cluster", device = ""} : () -> ()

  %outputs_8 =  "tf.Const"() {device = "", value = dense<1> : tensor<i32>} : () -> tensor<i32>
  %outputs_10 =  "tf.Const"() {device = "", value = dense<256> : tensor<i32>} : () -> tensor<i32>
  %outputs_12 =  "tf.Const"() {device = "", value = dense<0> : tensor<i32>} : () -> tensor<i32>

    func.return %0 : tensor<2x100xf32>
  }
  func.func @func_0_GPU_FLOAT(%arg0: tensor<100xf32>, %arg1: tensor<100xf32>, %arg2: tensor<100xf32>) -> tensor<100xf32> attributes {tac.cost = 4.000000e+01 : f32, tac.device = "GPU", tac.inference_type = "FLOAT", tac.interface_name = "func_0"} {

      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.
//

  %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>

    %1 = "tf.Relu"(%0) {device = "tpu0"} : (tensor<?xi32>) -> tensor<?xi32>
    %2 = "tf.Add"(%0, %1) {device = "tpu0"} : (tensor<?xi32>, tensor<?xi32>) -> tensor<?xi32>
    %3 = "tf.Relu"(%arg1) {device = "tpu1"} : (tensor<?xi32>) -> tensor<?xi32>
    %4 = "tf.Add"(%3, %arg1) {device = "tpu1"} : (tensor<?xi32>, tensor<?xi32>) -> tensor<?xi32>
    %5 = "tf.Relu"(%4) {device = "tpu0"} : (tensor<?xi32>) -> tensor<?xi32>

  %0 = "tf.Const"() {device = "/device:CPU:0", value = dense<2> : tensor<i32>} : () -> tensor<i32>
  %1 = "tf.ReadVariableOp"(%arg) {device = "/device:CPU:0", dtype = i32} : (tensor<!tf_type.resource<tensor<i32>>>) -> tensor<i32>
  %2 = "tf.Add"(%1, %0) {device = "/device:CPU:0"} : (tensor<i32>, tensor<i32>) -> tensor<i32>

  // CHECK: {{%.*}} = tfrt_fallback_async.executeop {{.*}} device("/device:CPU:0") "tf.MatMul"
  // CHECK-SAME: {T = f32, transpose_a = false, transpose_b = false}

  %dense_default_1 = "tf.Const"() {device = "/device:CPU:0", dtype = f32, value = dense<[]> : tensor<0xf32>} : () -> tensor<0xf32>
  %dense_keys = "tf.Const"() {device = "/device:CPU:0", dtype = !tf_type.string, value = dense<""> : tensor<2x!tf_type.string>} : () -> tensor<2x!tf_type.string>

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;
};