%cst = "tf.Const"() {value = dense<[0, 1, 2, 3]> : tensor<4xi32>} : () -> tensor<4xi32>
    %cst_0 = "tf.Const"() {value = dense<5.0> : tensor<f32>} : () -> tensor<f32>
    // Check that these constants are remained in constants.
    // CHECK: func private @while_cond
    // CHECK-DAG:  %[[CST:.*]]= "tf.Const"() <{value = dense<[0, 1, 2, 3]> : tensor<4xi32>}> : () -> tensor<4xi32>

    module attributes {tf.versions = {bad_consumers = [], min_consumer = 0 : i32, producer = 268 : i32}} {
      func.func @main() -> (tensor<2xi32>) {
        %cst = "tf.Const"() {value = dense<[524170, 523952]> : tensor<2xi32>} : () -> tensor<2xi32>
        return %cst : tensor<2xi32>
    }
  })";

  auto compilation_result = CompileMlirModule(
      kHasReturnValuesAndNoMetadataRetvals,

    %cst = "tf.Const"() <{value = dense<[0, 1]> : tensor<2xi32>}> {device = ""} : () -> tensor<2xi32>
    %cst_0 = "tf.Const"() <{value = dense<1.000000e+01> : tensor<f32>}> {device = ""} : () -> tensor<f32>
    %0 = "tf.Sum"(%arg0, %cst) <{keep_dims = false}> {device = ""} : (tensor<1x4xf32>, tensor<2xi32>) -> tensor<f32>
    %1 = "tf.Greater"(%0, %cst_0) {device = ""} : (tensor<f32>, tensor<f32>) -> tensor<i1>

func.func @merge_with_variant_type(%arg0: tensor<!tf_type.variant>, %arg1: tensor<!tf_type.variant<tensor<4xi32>>>) -> tensor<!tf_type.variant<tensor<8xf32>>> {
  %result = tf_executor.graph {

// CHECK: tf_executor.Merge{{.*}}(tensor<!tf_type.variant>, tensor<!tf_type.variant<tensor<4xi32>>>) -> (tensor<!tf_type.variant<tensor<8xf32>>>, tensor<i32>, !tf_executor.control)

func.func @QuantizeTransposeConv(%arg0: tensor<32x4x4x128xf32>, %arg1: tensor<4xi32>) -> tensor<1x32x42x128xf32> {
  %w = arith.constant dense<127.0> : tensor<1x32x42x128xf32>
  %b = arith.constant dense<0.0> : tensor<1x32x42x128xf32>

    %cst = "tf.Const"() <{value = dense<[0, 1]> : tensor<2xi32>}> {device = ""} : () -> tensor<2xi32>
    %cst_0 = "tf.Const"() <{value = dense<1.000000e+01> : tensor<f32>}> {device = ""} : () -> tensor<f32>
    %0 = "tf.Sum"(%arg0, %cst) <{keep_dims = false}> {device = ""} : (tensor<1x4xf32>, tensor<2xi32>) -> tensor<f32>
    %1 = "tf.Greater"(%0, %cst_0) {device = ""} : (tensor<f32>, tensor<f32>) -> tensor<i1>

# MLIR-SAME:  control_outputs = ""
# MLIR-SAME:  inputs = "input"
# MLIR-SAME:  outputs = "output"
# MLIR:         %[[shape:.*]] = arith.constant dense<[1, -1, 31]> : tensor<3xi32>
# MLIR:         %[[bias:.*]] = "tfl.pseudo_qconst"() <{qtype = tensor<186x!quant.uniform<i32:f32:0

# MLIR-SAME:  control_outputs = ""
# MLIR-SAME:  inputs = "input"
# MLIR-SAME:  outputs = "output"
# MLIR:         %[[shape:.*]] = arith.constant dense<[1, -1, 31]> : tensor<3xi32>
# MLIR:         %[[bias:.*]] = "tfl.pseudo_qconst"() <{qtype = tensor<186x!quant.uniform<i32:f32:0
# MLIR:         %[[weight:.*]] = "tfl.pseudo_qconst"() <{qtype = tensor<186x1x1x256x!quant.uniform<i4<-7:7>:f32:0, {2.2825599397931779,

// PerTensor-LABEL: QuantizeGatherWeightOnly
func.func @QuantizeGatherWeightOnly(%arg0: tensor<3xi32>) -> tensor<3x3x3x3xf32> {
  %w = arith.constant dense<1.270000e+02> : tensor<64x3x3x3xf32>
  %emb = "tfl.gather"(%w, %arg0) {axis = 0 : i32, batch_dims = 0 : i32} : (tensor<64x3x3x3xf32>, tensor<3xi32>) -> tensor<3x3x3x3xf32>

  %cst_0 = "tf.Const"() {value = dense<[-1, 10]> : tensor<2xi32>} : () -> tensor<2xi32>
  %1 = "tf.MatMul"(%arg0, %arg1) {
    transpose_a = false, transpose_b = true
  } : (tensor<1x10xf32>, tensor<10x10xf32>) -> tensor<*xf32>
  %2 = "tf.Reshape"(%1, %cst_0) : (tensor<*xf32>, tensor<2xi32>) -> tensor<*xf32>