-> (tensor<i64>, tensor<i32>) attributes {tf._tf_data_function = true, tf.signature.is_stateful} {
    %0:2 = "tf.JustPretend"(%arg2, %arg3, %arg4, %arg5) : (tensor<32xf32>, tensor<64xf64>, tensor<!tf_type.resource<tensor<64xf32>>>, tensor<!tf_type.resource<tensor<128xf32>>>) -> (tensor<i64>, tensor<i32>)
    func.return %0#0, %0#1: tensor<i64>, tensor<i32>

  %9 = mhlo.floor %8 : tensor<64xf32>
  %10 = mhlo.convert %9 : (tensor<64xf32>) -> tensor<64xi32>
  %11 = mhlo.compare  LT, %10, %1,  SIGNED : (tensor<64xi32>, tensor<64xi32>) -> tensor<64xi1>
  %12 = mhlo.add %10, %0 : tensor<64xi32>
  %13 = mhlo.select %11, %12, %10 : tensor<64xi1>, tensor<64xi32>
  %14 = mhlo.reshape %13 : (tensor<64xi32>) -> tensor<64x1xi32>

        tensor<20xf32>, tensor<20xf32>, tensor<20xf32>,
        tensor<20xf32>, tensor<20xf32>, tensor<20xf32>, tensor<20xf32>,
        tensor<20x20xf32>, none,
        tensor<1x20xf32>, tensor<1x20xf32>,
        none, none, none, none) -> tensor<1x28x20xf32>
    %1 = "quantfork.stats"(%0) {layerStats = dense<[-1.0, 2.0]> : tensor<2xf32>} : (tensor<1x28x20xf32>) -> tensor<1x28x20xf32>

  // CHECK: %[[VAL_4:.*]] = "tf.Identity"(%[[VAL_3]]) : (tensor<*xf32>) -> tensor<*xf32>
  // CHECK: return %[[VAL_4]]
  %3 = "tf.MatMul"(%arg1, %arg2) {transpose_a = false, transpose_b = false} : (tensor<8x32xf32>, tensor<32x64xf32>) -> tensor<*xf32>
  %4 = "tf.BiasAdd"(%3, %arg0) {data_format = "NHWC"} : (tensor<*xf32>, tensor<64xf32>) -> tensor<*xf32>
  %5 = "tf.Elu"(%4) : (tensor<*xf32>) -> tensor<*xf32>

  %0 = stablehlo.constant dense<2.000000e+00> : tensor<167x64xf32>
  %1 = stablehlo.constant dense<2.000000e+00> : tensor<64xf32>
  %2 = stablehlo.dot_general %arg0, %0, contracting_dims = [2] x [0], precision = [DEFAULT, DEFAULT] : (tensor<1x1x167xf32>, tensor<167x64xf32>) -> tensor<1x1x64xf32>
  %3 = stablehlo.broadcast_in_dim %1, dims = [2] : (tensor<64xf32>) -> tensor<1x1x64xf32>
  %4 = stablehlo.add %2, %3 : tensor<1x1x64xf32>
  func.return %4: tensor<1x1x64xf32>
}

  %0 = "tfl.unidirectional_sequence_lstm"(%arg0,...

  %1 = "tf.AddV2"(%arg0, %0) : (tensor<f32>, tensor<4xf32>) -> tensor<4xf32>
  %2 = "tf.AddV2"(%0, %arg0) : (tensor<4xf32>, tensor<f32>) -> tensor<4xf32>

  // If operand has the same shape as a result, we can fold it.
  %3 = "tf.AddV2"(%arg1, %0) : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32>
  %4 = "tf.AddV2"(%0, %arg1) : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32>

  // CHECK: %[[CONST:.*]] = "tf.Const"()

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

// -----

// CHECK-LABEL: fusedBatchNormV3_training_dynamic_unsupported1

    // CHECK: %[[MUL:.*]] = "tf.Mul"{{.*}} (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
    // CHECK: %[[UNKNOWN:.*]] = "tf.Unknown"(%[[MUL]], %[[MUL]]) : (tensor<1xf32>, tensor<1xf32>) -> tensor<*xf32>
    // CHECK: return %[[UNKNOWN]] : tensor<*xf32>
    %0 = "tf.Mul"(%arg0, %arg0) : (tensor<1xf32>, tensor<1xf32>) -> tensor<*xf32>
    %1 = "tf.Unknown"(%0, %0) : (tensor<*xf32>, tensor<*xf32>) -> tensor<*xf32>

  // expected-error @+1 {{requires mean to be a 1D float tensor}}
  %0:5 = "tf.FusedBatchNorm"(%arg0, %arg1, %arg2, %arg3, %arg4) {T = "tfdtype$DT_FLOAT", data_format = "NHWC", epsilon = 0.001 : f32, is_training = false} : (tensor<8x8x8x8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<?x8xf32>, tensor<8xf32>) -> (tensor<8x8x8x8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<?x8xf32>, tensor<8xf32>)