%2 = "stablehlo.concatenate"(%0, %1) {dimension = 0 : i64} : (tensor<?xf32>, tensor<?xf32>) -> tensor<?xf32>
  %3 = "quantfork.stats"(%2) {layerStats = dense<[-0.83811146, 2.4960899]> : tensor<2xf32>} : (tensor<?xf32>) -> tensor<?xf32>
  %4 = "quantfork.stats"(%arg2) {layerStats = dense<[-1.5726943, 1.07351148]> : tensor<2xf32>} : (tensor<?xf32>) -> tensor<?xf32>

// CHECK-DAG: %[[CAST_0:.*]] = "tf.Cast"(%[[ARG_0]]) <{Truncate = false}> : (tensor<2xbf16>) -> tensor<2xf32>
// CHECK-DAG: %[[CAST_1:.*]] = "tf.Cast"(%[[ARG_1]]) <{Truncate = false}> : (tensor<2xbf16>) -> tensor<2xf32>
// CHECK: %[[ADD:.*]] = "tf.AddV2"(%[[CAST_0]], %[[CAST_1]]) : (tensor<2xf32>, tensor<2xf32>) -> tensor<2xf32>

    "tf.Yield"(%2) : (tensor<i1>) -> ()
  }, {
  ^body(%barg0: tensor<*xf32>):
    %2 = "tf.SomeOp"(%barg0) : (tensor<*xf32>) -> tensor<*xf32>
    "tf.Yield"(%2) : (tensor<*xf32>) -> ()
  }) {is_stateless = false, shape_invariant} : (tensor<4xf32>) -> (tensor<*xf32>)

  func.return %0, %1 : tensor<*xf32>, tensor<*xf32>
}

// Test that both passes drop the shape_invariant attribute from
// While/WhileRegion ops within a cluster.

    %cst_0 = "tf.Const"() {value = dense<1.000000e+00> : tensor<2xf32>} : () -> tensor<2xf32>
    %var_0 = "tf.VarHandleOp"() {shared_name = "var_0"} : () -> tensor<!tf_type.resource<tensor<2xf32>>>
    "tf.AssignVariableOp"(%var_0, %cst_0) : (tensor<!tf_type.resource<tensor<2xf32>>>, tensor<2xf32>) -> ()
    %add_0 = "tf.Identity"(%cst_0) : (tensor<2xf32>) -> tensor<2xf32>

    %cst_1 = "tf.Const"() {value = dense<3.000000e+00> : tensor<2x3x3x512xf32>} : () -> tensor<2x3x3x512xf32>
    %0 = "tf.PartitionedCall"(%arg0, %cst_1) {_tfl_quant_trait = "fully_quantizable", config = "", config_proto = "", executor_type = "", f = @composite_depthwise_conv2d_fn} : (tensor<1x3x4x512xf32>, tensor<2x3x3x512xf32>) -> tensor<*xf32>

    %4 = "quantfork.stats"(%2) {layerStats = dense<[-2795.36523, 4609.57373]> : tensor<2xf32>} : (tensor<*xf32>) -> tensor<*xf32>
    %5 = "tf.BiasAdd"(%4, %3) {data_format = "NHWC", device = ""} : (tensor<*xf32>, tensor<2xf32>) -> tensor<*xf32>
    %6 = "quantfork.stats"(%5) {layerStats = dense<[-2788.31055, 4616.62842]> : tensor<2xf32>} : (tensor<*xf32>) -> tensor<*xf32>

    %3 = "tf.XlaCallModule"(%2, %1, %0) <{Sout = [#tf_type.shape<?x2>], module = "", version = 9 : i64}> {_entry_function = @composite_dot_general_with_bias_and_relu6_dynamic_fn_1, _original_entry_function = "composite_dot_general_with_bias_and_relu6_dynamic_fn_1", _tfl_quant_trait = "fully_quantizable"} : (tensor<?x2xf32>, tensor<2x2xf32>, tensor<2xf32>) -> tensor<?x2xf32>

  // CHECK-NEXT: }

  %0 = "tfl.pseudo_const" () {value = dense<1.0> : tensor<4xf32>} : () -> tensor<4xf32> loc("Const")
  %1 = "tfl.squared_difference"(%arg0, %0) {fused_activation_function = "NONE"} : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32> loc("squared_difference")
  %2 = "tfl.mul"(%arg0, %1) {fused_activation_function = "NONE"} : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32> loc("mul")

      %cst_1 = arith.constant dense<2.0> : tensor<3xf32>
      %0 = "tf.XlaCallModule"(%arg0, %cst_0, %cst_1) <{Sout = [#tf_type.shape<1x4x4x3>], module = "", version = 9 : i64}> {_entry_function = @composite_fn_1, _original_entry_function = "composite_fn_1", _tfl_quant_trait = "fully_quantizable"} : (tensor<1x4x4x3xf32>, tensor<3x1x1x3xf32>, tensor<3xf32>) -> tensor<1x4x4x3xf32>

  %biasadd = "tf.BiasAdd"(%conv, %dq_bias) {data_format = "NHWC", device = ""} : (tensor<*xf32>, tensor<2xf32>) -> tensor<*xf32>
  %res = "tf.Relu6"(%biasadd) : (tensor<*xf32>) -> tensor<*xf32>
  %q_res = "quantfork.qcast"(%res) : (tensor<*xf32>) -> tensor<*x!quant.uniform<i8:f32, 0.023529411764705882:-128>>