func.func @tf2xla_fallback_op_approx_top_k(%arg0: tensor<16xf32>) -> (tensor<?xf32>, tensor<?xi32>) {
  %0:2 = "tf_device.cluster"() ({
    // CHECK: tf.ApproxTopK
    // CHECK-NOT: _xla_outside_compilation
    %1:2 = "tf.ApproxTopK"(%arg0) {k = 2} : (tensor<16xf32>) -> (tensor<?xf32>, tensor<?xi32>)
    tf_device.return %1#0, %1#1 : tensor<?xf32>, tensor<?xi32>

  // CHECK-NEXT: %[[UPDATE_SLICE:.*]] = "tf.Reshape"(%[[PUSHVAL]], %[[UPDATE_SHAPE]]) : (tensor<f32>, tensor<1xi32>) -> tensor<1xf32>
  // CHECK-NEXT: %[[UPDATE:.*]] = "tf.XlaDynamicUpdateSlice"(%[[BROADCAST]], %[[UPDATE_SLICE]], %[[ZERO]]) : (tensor<10xf32>, tensor<1xf32>, tensor<1xi32>) -> tensor<10xf32>

  // CHECK-NEXT: %[[CONST1:.*]] = "tf.Const"() <{value = dense<1> : tensor<1xi32>}> : () -> tensor<1xi32>
  // CHECK-NEXT: %[[NEW_SIZE:.*]] = "tf.AddV2"(%[[READ_SIZE]], %[[CONST1]]) : (tensor<1xi32>, tensor<1xi32>) -> tensor<1xi32>
  // CHECK-NEXT: "tf.AssignVariableOp"(%[[SIZE]], %[[NEW_SIZE]]) : (tensor<!tf_type.resource<tensor<1xi32>>>, tensor<1xi32>) -> ()

    %4 = "tf.LayerNorm"(%a1, %a2, %a3, %a4) {_tfl_quant_trait = "fully_quantizable", device = ""} : (tensor<128x128xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xi32>) -> tensor<128x128xf32>
   "tfl.yield"(%4) : (tensor<128x128xf32>) -> ()
  }) {_tfl_quant_trait = "fully_quantizable", device = ""} : (tensor<128x128xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xi32>) -> tensor<128x128xf32>

  attributes {tf_saved_model.exported_names = ["assign_from_global"]} {
    // CHECK: [[name:%.*]] = "tf.VarHandleOp"
    // CHECK: "tf.AssignVariableOp"([[name]]
    %0 = "tf.Const"() {type = tensor<10xf32>, value = dense<[0.,10.,2.,3.,4.,5.,6.,7.,8.,9.]> : tensor<10xf32>} : () -> tensor<10xf32>
    "tf.AssignVariableOp"(%arg0, %0) : (tensor<!tf_type.resource<tensor<10xf32>>>, tensor<10xf32>) -> ()

module {
  func.func @simpleTest(%arg0: tensor<1xf32>, %arg1: tensor<1xf32>, %arg2: tensor<1xf32>, %arg3: tensor<1xf32>) -> tensor<2x1xf32> {
    %0 = func.call @func_0_GPU_FLOAT(%arg0, %arg1, %arg2) {tac.interface_name = "func_0"} : (tensor<1xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
    %1 = func.call @func_1_GPU_FLOAT(%arg0, %arg3) {tac.interface_name = "func_1"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>

  %2 = "tfl.add"(%arg0, %arg3) {tac.device = "GPU", fused_activation_function = "RELU6", tac.inference_type = "FLOAT"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
  %3 = "tfl.pack"(%1, %2) {tac.device = "CPU", tac.inference_type = "FLOAT", axis = 0 : i32, values_count = 2 : i32} : (tensor<1xf32>, tensor<1xf32>) -> tensor<2x1xf32>
  return %3 : tensor<2x1xf32>
}
```

func.func @multiply(%arg0: tensor<1xf32>, %arg1: tensor<1xf32>) -> tensor<1xf32> {
  %0 = "vhlo.multiply_v1"(%arg0, %arg1) : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
  func.return %0 : tensor<1xf32>
}

// CHECK:func.func private @multiply(%arg0: tensor<1xf32>, %arg1: tensor<1xf32>) -> tensor<1xf32> {
// CHECK: %0 = "vhlo.multiply_v1"(%arg0, %arg1) : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
// CHECK: return %0 : tensor<1xf32>
// CHECK:}

func.func @add(%arg0: tensor<1xf32>, %arg1: tensor<1xf32>) -> tensor<1xf32> {
  %0 = "tf.Add"(%arg0, %arg1) : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
  func.return %0: tensor<1xf32>
// CHECK: %[[CUSTOM_0:.*]] = "tfl.custom"(%arg0, %arg1) <{custom_code = "FlexAdd", custom_option = #tfl<const_bytes : "0x03416464001412034164641A001A002A070A015412023001320000021B171414042801">}> : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>

    func.return %0 : tensor<?xf32>
  }
}

// -----

module attributes {tf_saved_model.semantics} {

  "tf_saved_model.global_tensor"() { is_mutable, sym_name = "v", type = tensor<?xf32>, value = dense<1.> : tensor<1xf32> } : () -> ()
  // expected-error@+1 {{bound input with type 'tensor<f32>' expected to have type 'tensor<!tf_type.resource<tensor<?xf32>>>'}}