func.func @batchNormInference_2D_inner_features(
    %x: tensor<4x256xf32>, %scale: tensor<256xf32>, %offset: tensor<256xf32>,
    %mean: tensor<256xf32>, %variance: tensor<256xf32>)
    -> (tensor<4x256xf32>) {
  // CHECK-DAG: %[[EPS_BCAST:.+]] = mhlo.constant dense<1.001000e-05> : tensor<256xf32>
  // CHECK-DAG: %[[VARIANCE_EPS:.+]] = mhlo.add %[[VARIANCE]], %[[EPS_BCAST]] : tensor<256xf32>

  %1 = "tf.BiasAdd"(%0, %cst_0) {data_format = "NHWC"} : (tensor<?x?x?x256xf32>, tensor<256xf32>) -> tensor<?x?x?x256xf32>
  %2 = "tf.Mul"(%1, %cst_1) : (tensor<?x?x?x256xf32>, tensor<256xf32>) -> tensor<?x?x?x256xf32>
  func.return %2 : tensor<?x?x?x256xf32>
}
// CHECK: func @conv2d_with_large_weight_and_mul

// CHECK-NEXT:      %17 = mhlo.dynamic_reshape %arg1, %16 : (tensor<4x?x256xf32>, tensor<3xi32>) -> tensor<4x?x256xf32>
// CHECK-NEXT:      %18 = "tfl.batch_matmul"(%8, %17) <{adj_x = false, adj_y = false, asymmetric_quantize_inputs = false}> : (tensor<4x4x?xf32>, tensor<4x?x256xf32>) -> tensor<4x4x256xf32>
// CHECK-NEXT:      %19 = mhlo.reshape %18 : (tensor<4x4x256xf32>) -> tensor<4x4x256xf32>

    %3 = stablehlo.concatenate %2, %1, dim = 0 : (tensor<1x3xf32>, tensor<1x3xf32>) -> tensor<2x3xf32>
    return %3 : tensor<2x3xf32>
  }
  func.func private @composite_dot_general_fn_1(%arg0: tensor<1x2xf32>, %arg1: tensor<2x3xf32>) -> tensor<1x3xf32> attributes {_from_xla_call_module, tf_quant.composite_function} {

  %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>

  %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>

  %1:4 = "tfl.split"(%cst, %0) {num_splits = 4 : i32} : (tensor<i32>, tensor<4xf32>) -> (tensor<2xf32>, tensor<2xf32>,tensor<2xf32>, tensor<2xf32>)
  %2 = "tfl.quantize"(%1#0) {qtype = tensor<2x!quant.uniform<u8:f32, 1.0>>} : (tensor<2xf32>) -> tensor<2x!quant.uniform<u8:f32, 1.0>>
  %3 = "tfl.quantize"(%1#1) {qtype = tensor<2x!quant.uniform<u8:f32, 1.0>>} : (tensor<2xf32>) -> tensor<2x!quant.uniform<u8:f32, 1.0>>

  // CHECK-SAME: %[[ARG1:.*]]: tensor<2x4xf32>
  func.func private @composite_and_slice(%arg0: tensor<3x2xf32>, %arg1: tensor<2x4xf32>) -> tensor<2x2xf32> {
    // CHECK: %[[Q1:.*]] = "quantfork.qcast"(%[[ARG0]]) {volatile} : (tensor<3x2xf32>) -> tensor<3x2x!quant.uniform<i8:f32, 5.000000e-03>>

// CHECK: [[UNPACK:%.*]]:3 = "tf.Unpack"(%arg0) <{axis = 0 : i64}> : (tensor<3x2x2xf32>) -> (tensor<2x2xf32>, tensor<2x2xf32>, tensor<2x2xf32>)
// CHECK: [[SCALAR_ZERO:%.*]] = arith.constant dense<0> : tensor<i32>
// CHECK: [[CONCAT:%.*]] = "tf.Concat"([[SCALAR_ZERO]], [[UNPACK]]#0, [[UNPACK]]#1, [[UNPACK]]#2) : (tensor<i32>, tensor<2x2xf32>, tensor<2x2xf32>, tensor<2x2xf32>) -> tensor<?x2xf32>
// CHECK: [[LENGTHS:%.*]] = arith.constant dense<0> : tensor<0xi64>

    version = 5 : i64
  } : (tensor<2x2xf32>, tensor<2x3xf32>) -> tensor<2x3xf32>
  %2 = stablehlo.add %cst, %0 : tensor<2x3xf32>
  return %2 : tensor<2x3xf32>
}

// CHECK-LABEL: func.func @no_qdq_for_multiple_users
// CHECK-NOT: quantfork.qcast