```mlir
func @main(%arg0: tensor<8x4xf32>) {
  %elem_shape = "tf.Const"() {value = dense<[8, 4]> : tensor<2xi32>} : () -> tensor<2xi32>
  %max_size = "tf.Const"() {value = dense<10> : tensor<i32>} : () -> tensor<i32>
  %tl = "tf.EmptyTensorList"(%elem_shape, %max_size) : (tensor<2xi32>, tensor<i32>) -> tensor<!tf_type.variant<tensor<8x4xf32>>>

func.func @testNMSV3ToNMSV4(%arg0: tensor<3x4xf32>, %arg1: tensor<3xf32>, %arg2: tensor<f32>, %arg3: tensor<f32>) -> tensor<2xi32> {
  %max_size = arith.constant dense<2> : tensor<i32>
  // CHECK: "tf.NonMaxSuppressionV4"
  // CHECK-SAME: device = "/job:localhost/replica:0/task:0/device:GPU:0"

    public LoadingValueReference() {
      this(LocalCache.<K, V>unset());
    }

    /*
     * TODO(cpovirk): Consider making this implementation closer to the mainline implementation.
     * (The difference was introduced as part of Java-8-specific changes in cl/132882204, but we
     * could probably make *some* of those changes here in the backport, too.)
     */