%2 = func.call @func_2_CPU_FLOAT(%0, %1) {tac.interface_name = "func_2"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<2x1xf32>
    func.return %2 : tensor<2x1xf32>
  }

  func.func private @func_2_CPU_FLOAT(%arg0: tensor<1xf32>, %arg1: tensor<1xf32>) -> tensor<2x1xf32> attributes {tac.device = "CPU", tac.inference_type = "FLOAT", tac.interface_name = "func_2"} {

func.func @torch_index_select(%arg0: tensor<2x1xf32>, %arg1: tensor<2xi32>) -> tensor<2x1xf32> {
  %0 = "mhlo.torch_index_select"(%arg0, %arg1) {
    batch_dims = 0 : i64, dim = 0 : i64
  } : (tensor<2x1xf32>, tensor<2xi32>) -> tensor<2x1xf32>
  func.return %0 : tensor<2x1xf32>
}

// CHECK-LABEL:   func @lowered_cumsum(
// CHECK-SAME:      %[[VAL_0:.*]]: tensor<4x12xf32>) -> tensor<4x12xf32> {

  %3 = "tfl.dequantize"(%2) : (tensor<2x2x!quant.uniform<u8:f32, 1.0>>) -> tensor<2x2xf32>
  func.return %3 : tensor<2x2xf32>

// CHECK-NEXT: %[[Q:.*]] = "tfl.quantize"(%arg0) <{qtype = tensor<1x2x!quant.uniform<u8:f32, 1.000000e+00>>}> : (tensor<1x2xf32>) -> tensor<1x2x!quant.uniform<u8:f32, 1.000000e+00>>

func.func @add_with_activation_transpose_rank_two(%arg0: tensor<1x2xf32>) -> tensor<2x1xf32> {
  %0 = stablehlo.constant dense<2.000000e+00> : tensor<2x1xf32>
  %1 = stablehlo.transpose %arg0, dims = [1, 0] : (tensor<1x2xf32>) -> tensor<2x1xf32>
  %2 = stablehlo.add %1, %0 : tensor<2x1xf32>
  return %2 : tensor<2x1xf32>
}
// CHECK: %[[TRANSPOSE_0:.+]] = stablehlo.transpose

func.func @uniform_quantize_and_dequantize_per_axis(%arg0 : tensor<2x2xf32>) -> tensor<2x2xf32> {
  %scales = "tf.Const"() { value = dense<[1.0, 2.0]> : tensor<2xf32> } : () -> tensor<2xf32>
  %zps = "tf.Const"() { value = dense<[3, 4]> : tensor<2xi32> } : () -> tensor<2xi32>

  // CHECK: %[[QUANTIZE:.*]] = mhlo.uniform_quantize %arg0 : (tensor<2x2xf32>) -> tensor<2x2x!quant.uniform<i8:f32:0, {1.000000e+00:3,2.000000e+00:4}>>

func.func @pack(%arg0: tensor<1xf32>, %arg1: tensor<1xf32>) -> tensor<2x1xf32> {
  %0 = "tfl.pack"(%arg0, %arg1) {axis = 0 : i32, values_count = 2 : i32} : (tensor<1xf32>, tensor<1xf32>) -> tensor<2x1xf32>
  func.return %0 : tensor<2x1xf32>
}

// CHECK:   func @pack(%[[VAL_0:.*]]: tensor<1xf32>, %[[VAL_1:.*]]: tensor<1xf32>) -> tensor<2x1xf32> {
// CHECK-DAG:       %[[VAL_2:.*]] = "tfl.pseudo_const"{{.*}}dense<1> : tensor<4xi32>

func.func @mean_4d_no_keepdim(%arg0: tensor<1x48x48x512xf32>) -> tensor<1x512xf32> {
  %cst = arith.constant dense<[1, 2]> : tensor<2xi32>
  %0 = "tfl.mean"(%arg0, %cst) {keep_dims = false} : (tensor<1x48x48x512xf32>, tensor<2xi32>) -> tensor<1x512xf32>
  func.return %0 : tensor<1x512xf32>
}

// CHECK:       func @mean_4d_no_keepdim([[VAL_0:%.*]]: tensor<1x48x48x512xf32>) -> tensor<1x512xf32> {

  %3 = "tfl.pack"(%1, %2) {axis = 0 : i32, per_device_costs = {CPU = 2.0 : f32, GPU = -1.0 : f32}, values_count = 2 : i32, tac.device = "CPU"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<2x1xf32>
  func.return %3 : tensor<2x1xf32>
})";
  const std::string kExpectedFB = CreateRuntimeMetadata();
  mlir::DialectRegistry registry;
  registry.insert<mlir::TFL::TensorFlowLiteDialect, mlir::arith::ArithDialect,

// CHECK: %0 = "tfl.pseudo_const"() <{value = dense<2> : tensor<2xi32>}> : () -> tensor<2xi32>
// CHECK: %1 = "tfl.reshape"(%arg0, %0) : (tensor<1x4xf32>, tensor<2xi32>) -> tensor<2x2xf32>

{
  "version": 3,
  "operator_codes": [
    {
      "builtin_code": "RESHAPE"
    }
  ],
  "subgraphs": [
    {
      "tensors": [
        {
          "shape": [1, 4],

    %4:4 = "tf.CustomAggregator"(%3) {calibration_method = 1 : i32, id = "1", max_percentile = 0.000000e+00 : f32, min_percentile = 0.000000e+00 : f32, num_bins = 0 : i32} : (tensor<3x11xf32>) -> (tensor<3x11xf32>, tensor<f32>, tensor<f32>, tensor<*xi64>)
    %5 = stablehlo.add %4#0, %1 : tensor<3x11xf32>
    %6 = stablehlo.multiply %5, %1 : tensor<3x11xf32>