%0 = tf_executor.graph {
    %outputs, %control = tf_executor.island {
      %1 = "tf.Const"() {value = dense<1> : tensor<i32>} : () -> tensor<i32>
      "tf.AssignVariableOp"(%arg0, %1) : (tensor<*x!tf_type.resource<tensor<i32>>>, tensor<i32>) -> ()
      tf_executor.yield %1 : tensor<i32>
    }
    tf_executor.fetch %outputs : tensor<i32>
  }
  func.return %0 : tensor<i32>

    "tfl.yield"(%13, %arg2, %11, %arg4, %12) : (tensor<i32>, tensor<i32>, tensor<i32>, tensor<!tf_type.resource>, tensor<f32>) -> ()
  }) {is_stateless = false} : (tensor<i32>, tensor<i32>, tensor<i32>, tensor<!tf_type.resource>, tensor<f32>) -> (tensor<i32>, tensor<i32>, tensor<i32>, tensor<!tf_type.resource>, tensor<f32>)
  func.return %5#3 : tensor<!tf_type.resource>

// CHECK: "tfl.while"

  %conv = "tfl.conv_2d"(%0, %w, %b) {dilation_h_factor = 1 : i32, dilation_w_factor = 1 : i32, fused_activation_function = "NONE", padding = "SAME", stride_h = 2 : i32, stride_w = 2 : i32} : (tensor<1x224x224x3xf32>, tensor<64x3x3x3xf32>, tensor<64xf32>) -> tensor<1x112x112x64xf32>

          tf_device.return %4 : tensor<i64>
        }) : () -> (tensor<i32>, tensor<i64>)
        tf_device.return %3#0, %3#1 : tensor<i32>, tensor<i64>
      }
      tf_executor.yield %2#0, %2#1, %2#2, %2#3 : tensor<i32>, tensor<i32>, tensor<i64>, tensor<i64>
    }
    tf_executor.fetch %1#0, %1#1, %1#2, %1#3 : tensor<i32>, tensor<i32>, tensor<i64>, tensor<i64>
  }
  func.return
}

  %2 = "tfl.pseudo_qconst"() {qtype = tensor<32x!quant.uniform<i32:f32, 1.7052092479439231E-4>>, value = dense<0> : tensor<32xi32>} : () -> tensor<32x!quant.uniform<i32:f32, 1.7052092479439231E-4>>

  %input_zp = "tf.Const"() { value = dense<-10> : tensor<i32> } : () -> tensor<i32>
  %filter_scale = "tf.Const"() { value = dense<0.5> : tensor<f32> } : () -> tensor<f32>
  %filter_zp = "tf.Const"() { value = dense<0> : tensor<i32> } : () -> tensor<i32>
  %accum_scale = "tf.Const"() { value = dense<1.0> : tensor<f32> } : () -> tensor<f32>
  %accum_zp = "tf.Const"() { value = dense<0> : tensor<i32> } : () -> tensor<i32>

func.func @stateful_composite_op_control_else(%arg0: tensor<*x!tf_type.resource<tensor<i32>>>) -> tensor<i32> {
  %0 = tf_executor.graph {
    %outputs, %control = tf_executor.island {
      %1 = "tf.Const"() {value = dense<1> : tensor<i32>} : () -> tensor<i32>
      "tf.AssignVariableOp"(%arg0, %1) : (tensor<*x!tf_type.resource<tensor<i32>>>, tensor<i32>) -> ()
      tf_executor.yield %1 : tensor<i32>
    }
    tf_executor.fetch %outputs : tensor<i32>
  }

  %0 = "tfl.conv_2d"(%arg0, %arg1, %arg2) {tac.device = "GPU", tac.inference_type = "FLOAT", dilation_h_factor = 1 : i32, dilation_w_factor = 1 : i32, fused_activation_function = "NONE", padding = "VALID", stride_h = 1 : i32, stride_w = 1 : i32} : (tensor<256x32x32x3xf32>, tensor<16x3x3x3xf32>, tensor<16xf32>) -> tensor<256x30x30x16xf32>
  func.return %0 : tensor<256x30x30x16xf32>
}

      func.func @main() -> tensor<!tf_type.variant<tensor<64x1xbf16>>> {
      // unknown rank
      %elem_shape = "tf.Const"() <{value = dense<-1> : tensor<i32>}> {device = "/job:localhost/replica:0/task:0/device:CPU:0"} : () -> tensor<i32>
      // zero reserved elements

// CKECK-SAME: <{calibration_method = 5 : i32, id = "0", num_bins = 32 : i32, max_percentile = 0.000000e+00 : f32, min_percentile = 0.000000e+00 : f32}> : (tensor<1x3x4x3xf32>) -> (tensor<1x3x4x3xf32>, tensor<f32>, tensor<f32>, tensor<512xi64>)
// CHECK: %[[CUSTOM_AGGREGATOR_1:.*]], %[[MIN_1:.*]], %[[MAX_1:.*]], %[[HISTOGRAM_1:.*]] = "tf.CustomAggregator"