%min = arith.constant dense<0.0> : tensor<16xf32>
  %max = arith.constant dense<15.0> : tensor<16xf32>
  %mini = "tf.Identity"(%min) : (tensor<16xf32>) -> tensor<16xf32>
  %maxi = "tf.Identity"(%max) : (tensor<16xf32>) -> tensor<16xf32>
  %fq = "tf.FakeQuantWithMinMaxVarsPerChannel"(%in, %mini, %maxi) {num_bits = 3, narrow_range = false} : (tensor<3x3x3x16xf32>, tensor<16xf32>, tensor<16xf32>) -> tensor<3x3x3x16xf32>

  %min = arith.constant dense<0.0> : tensor<16xf32>
  %max = arith.constant dense<255.0> : tensor<16xf32>
  %mini = "tf.Identity"(%min) : (tensor<16xf32>) -> tensor<16xf32>
  %maxi = "tf.Identity"(%max) : (tensor<16xf32>) -> tensor<16xf32>
  %fq = "tf.FakeQuantWithMinMaxVarsPerChannel"(%in, %mini, %maxi) {num_bits = 5, narrow_range = false} : (tensor<3x3x3x16xf32>, tensor<16xf32>, tensor<16xf32>) -> tensor<3x3x3x16xf32>

  %cst0 = mhlo.constant dense<[1, 2, 3, 4]> : tensor<4xi32>
  // CHECK: %[[BROADCAST:.*]] = "mhlo.broadcast_in_dim"(%[[CONST]]) <{broadcast_dimensions = dense<3> : tensor<1xi64>}> : (tensor<4xi32>) -> tensor<1x1x2x4xi32>
  %0 = "mhlo.broadcast_in_dim"(%cst0) <{broadcast_dimensions = dense<3> : tensor<1xi64>}> : (tensor<4xi32>) -> tensor<1x1x2x4xi32>

// CHECK:           return %[[VAL_2]], %[[VAL_3]] : tensor<4xf32>, tensor<4xf32>
// CHECK:         }
func.func @broadcast_atan2(%arg0: tensor<1xf32>, %arg1: tensor<4xf32>) -> (tensor<4xf32>, tensor<4xf32>) {
  %0 = "mhlo.broadcast_in_dim"(%arg0) <{broadcast_dimensions = dense<[0]> : tensor<1xi64>}> : (tensor<1xf32>) -> tensor<4xf32>
  %1 = mhlo.atan2 %0, %arg1 : tensor<4xf32>
  %2 = mhlo.atan2 %arg1, %0 : tensor<4xf32>

module{
  func.func @main(%arg0: tensor<10xf32>) -> tensor<10xf32> {
    return %arg0 : tensor<10xf32>
  }
  func.func @func1(%arg0: tensor<10xf32>, %arg1: tensor<10xf32>) -> tensor<10xf32> {
    %0 = stablehlo.add %arg0, %arg1 : tensor<10xf32>
    %1 = stablehlo.add %arg0, %arg1 : tensor<10xf32>
    return %0 : tensor<10xf32>
  }
})mlir";

    %1 = arith.constant dense<[[2.3, 3.4, 4.5, 5.5]]> : tensor<1x4xf32>
    %2 = "tf.Add"(%0, %1) : (tensor<1x4xf32>, tensor<1x4xf32>) -> tensor<1x4xf32>
    %3 = tensor.cast %2 : tensor<1x4xf32> to tensor<1x?xf32>
    func.return %3 : tensor<1x?xf32>
}

module {
  func.func @wrap_composite_func(%arg0: tensor<*xf32>, %arg1: tensor<*xf32>) -> tensor<*xf32> {
    %0 = "tf.PartitionedCall"(%arg0, %arg1) <{f = @composite_conv2d_with_relu6_fn}> {_tfl_quant_trait = "fully_quantizable"}
          : (tensor<*xf32>, tensor<*xf32>) -> tensor<*xf32>
    func.return %0 : tensor<*xf32>
  }

  func.func @no_composite_func(%arg0: tensor<*xf32>, %arg1: tensor<*xf32>) -> tensor<*xf32> {

// PerChannel-DAG: %[[scale2:.*]] = "tf.Const"() <{value = dense<{{[0-9\.Ee\+\-]+}}> : tensor<6xf32>}> : () -> tensor<6xf32>
// PerChannel-DAG: %[[zp1:.*]] = "tf.Const"() <{value = dense<{{[0-9]+}}> : tensor<3xi32>}> : () -> tensor<3xi32>
// PerChannel-DAG: %[[zp2:.*]] = "tf.Const"() <{value = dense<{{[0-9]+}}> : tensor<6xi32>}> : () -> tensor<6xi32>

    quantization_max_val = 127 : i64
  } : (tensor<10xf32>, tensor<f32>, tensor<i32>) -> tensor<10x!tf_type.qint8>
  %1 = "tf.UniformDequantize"(%0, %scale, %zp) {
    quantization_axis = -1 : i64,
    quantization_min_val = -128 : i64,
    quantization_max_val = 127 : i64
  } : (tensor<10x!tf_type.qint8>, tensor<f32>, tensor<i32>) -> tensor<10xf32>
  return %1 : tensor<10xf32>
})mlir";

  %0 = stablehlo.constant dense<2.000000e+00> : tensor<4x2x3x3xf32>  // weight
  %1 = stablehlo.constant dense<3.000000e+00> : tensor<4xf32>  // bias
  %2 = stablehlo.constant dense<0.000000e+00> : tensor<1x4x5x5xf32>  // relu
  %3 = stablehlo.broadcast_in_dim %1, dims = [1] : (tensor<4xf32>) -> tensor<1x4x5x5xf32>