%cmp = "tf.Less"(%read, %cst) : (tensor<32xf32>, tensor<32xf32>) -> tensor<32xi1>
            %dims = arith.constant dense<0> : tensor<1xi32>
            %reduce = "tf.All"(%cmp, %dims) {keep_dims = false} : (tensor<32xi1>, tensor<1xi32>) -> tensor<i1>
            "tf.Yield"(%reduce) : (tensor<i1>) -> ()
        }, {
          ^bb0(%barg0: tensor<i32>, %barg1: !tf_res, %barg2: !tf_res):

    %cst = "tf.Const"() {value = dense<-1> : tensor<i64>} : () -> tensor<i64>
    %cst_0 = "tf.Const"() {value = dense<[1, 0]> : tensor<2xi32>} : () -> tensor<2xi32>
    %cst_1 = "tf.Const"() {value = dense<1.000000e+00> : tensor<f32>} : () -> tensor<f32>
    %cst_2 = "tf.Const"() {value = dense<0.00235294132> : tensor<f32>} : () -> tensor<f32>

//          (tensor<?x3x4x2xf32>) -> tensor<i32>
// %5 = stablehlo.reshape %4 :
//          (tensor<i32>) -> tensor<1xi32>
// %6 = stablehlo.concatenate %5, %0, %1, %2, dim = 0 :
//          (tensor<1xi32>, tensor<1xi32>, tensor<1xi32>, tensor<1xi32>)
//            -> tensor<4xi32>
// %7 = stablehlo.dynamic_broadcast_in_dim %arg2, %6
// %8 = stablehlo.add %3, %7
// ```
//

    %0 = "tfl.pseudo_const"() {value = dense<0.962260901> : tensor<1xf32>} : () -> tensor<1xf32>
    %1 = func.call @func_0_GPU_FLOAT(%arg0, %0) {tac.device = "GPU", tac.inference_type = "FLOAT", tac.interface_name = "func_0"} : (tensor<1x200x200x200xf32>, tensor<1xf32>) -> tensor<1x200x200x200xf32>
    %2 = "tfl.pseudo_const"() {value = dense<0.895973444> : tensor<1xf32>} : () -> tensor<1xf32>

    %cst_1 = "tf.Const"() {value = dense<[2, 1024]> : tensor<2xi32>} : () -> tensor<2xi32>
    // Original identity preserved.
    %cst_identity = "tf.Identity"(%cst_0) {device = ""} : (tensor<1024x2xf32>) -> tensor<1024x2xf32>
    %0 = "tf.Reshape"(%cst_identity, %cst_1) : (tensor<1024x2xf32>, tensor<2xi32>) -> tensor<2x1024xf32>

    %s0 = "tf.Const"() {value = dense<[501, 1, 32, 1280]> : tensor<4xi32>} : () -> tensor<4xi32>
    %s1 = "tf.Const"() {value = dense<[  1, 1,  1, 1280]> : tensor<4xi32>} : () -> tensor<4xi32>

    // CHECK: "tf.BroadcastGradientArgs"
    // CHECK-NOT: _xla_outside_compilation
    %r0, %r1 = "tf.BroadcastGradientArgs"(%s0, %s1) {} : (tensor<4xi32>, tensor<4xi32>) -> (tensor<1xi32>, tensor<3xi32>)
    tf_device.return

  %1 = "tfl.pseudo_const"() {value = dense<0> : tensor<3xi32>} : () -> tensor<3xi32>
  %2 = "tfl.pseudo_const"() {value = dense<[0, 2, 0]> : tensor<3xi32>} : () -> tensor<3xi32>
  %3 = "tfl.pseudo_const"() {value = dense<1> : tensor<3xi32>} : () -> tensor<3xi32>

  %4 = "tfl.var_handle"() {container = "", shared_name = "read_assign/states0"} : () -> tensor<!tf_type.resource>

  %0 = "tfr.cast"(%arg0) : (tensor<*xf32>) -> !tfr.tensor
  %1 = "tfr.cast"(%0) : (!tfr.tensor) -> tensor<*xi32>
  %2 = "tfr.cast"(%0) : (!tfr.tensor) -> tensor<2xi32>
  func.return %1, %2 : tensor<*xi32>, tensor<2xi32>

// CHECK: %[[tf_cast_unranked:.*]] = "tf.Cast"(%arg0) <{Truncate = false}> : (tensor<*xf32>) -> tensor<*xi32>

// 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>
// PerChannel: %[[out_1:.*]] = "tf.PartitionedCall"(%arg0, %[[q_w1]], %[[scale1]], %[[zp1]]) <{config = "", config_proto = "", executor_type = "",

func.func @add(%arg0: tensor<1xf32>, %arg1: tensor<1xf32>) -> tensor<1xf32> {
  %0 = "tf.Add"(%arg0, %arg1) : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
  func.return %0: tensor<1xf32>
// CHECK: %[[CUSTOM_0:.*]] = "tfl.custom"(%arg0, %arg1) <{custom_code = "FlexAdd", custom_option = #tfl<const_bytes : "0x03416464001412034164641A001A002A070A015412023001320000021B171414042801">}> : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>