%recurrent_input = "tfl.pseudo_const"() {value = dense<0.000000e+00> : tensor<1x3xf32>} : () -> tensor<1x3xf32>
  %recurrent_stats = "quantfork.stats"(%recurrent_input) {layerStats = dense<[0.0, 1.0]> : tensor<2xf32>} : (tensor<1x3xf32>) -> tensor<1x3xf32>
  %cell_input = "tfl.pseudo_const"() {value = dense<1.000000e+00> : tensor<1x3xf32>} : () -> tensor<1x3xf32>

func.func @dequantize_pushdown(%arg0: tensor<1000x1000x!quant.uniform<i8:f32, 7.812500e-03>>, %arg1: tensor<1x1xi32>) -> tensor<1x1x1000xf32> {
  %0 = "tfl.dequantize"(%arg0) : (tensor<1000x1000x!quant.uniform<i8:f32, 7.812500e-03>>) -> tensor<1000x1000xf32>
  %1 = "tfl.gather"(%0, %arg1) {axis = 0 : i32, batch_dims = 0 : i32}: (tensor<1000x1000xf32>, tensor<1x1xi32>) -> tensor<1x1x1000xf32>
  func.return %1 : tensor<1x1x1000xf32>

// CHECK-NEXT: tfl.gather

// CHECK-NEXT:    "tf.IfrtCall"(%arg0, [[ARRAYKEY]]) <{program_id = 6515870160938153680 : i64, variable_arg_indices = [1 : i32]}> {__tpu_compile_metadata_text = "retvals { sharding { } }"} : (tensor<1x3xf32>, tensor<!tf_type.string>) -> tensor<1x1xf32>
// CHECK-NEXT:    return
//
 func.func @serving_default(%arg0: tensor<1x3xf32>) -> tensor<1x1xf32> {

func.func @main(%arg0: tensor<1x4xf32>) -> tensor<1x3xf32> {
  %0 = stablehlo.constant dense<1.0> : tensor<4x3xf32>
  %1 = stablehlo.dot_general %arg0, %0, contracting_dims = [1] x [0], precision = [DEFAULT, DEFAULT] : (tensor<1x4xf32>, tensor<4x3xf32>) -> tensor<1x3xf32>
  return %1 : tensor<1x3xf32>
}
// CHECK: @main(%[[ARG_0:.+]]: tensor<1x4xf32>) -> tensor<1x3xf32>

  // CHECK: device = "cpu"
  %2 = "tf.Relu"(%1) {T = f32, _output_shapes = ["tfshape$dim { size: 3 } dim { size: 3 }"], device = "cpu"} : (tensor<3x3xf32>) -> tensor<3x3xf32>
  // CHECK: device = "xpu"
  %3 = "tf.Relu"(%2) {T = f32, _output_shapes = ["tfshape$dim { size: 3 } dim { size: 3 }"]} : (tensor<3x3xf32>) -> tensor<3x3xf32>
  func.return %3 : tensor<3x3xf32>
}

  %0 = "tfl.pseudo_qconst"() { qtype = tensor<3x3x!quant.uniform<i32:f32:1, {1.0, 0.5:1, 0.25:1}>>, value = dense<1> : tensor<3x3xi32>} : () -> tensor<3x3x!quant.uniform<i32:f32:1, {1.0, 0.5:1, 0.25:1}>>

module {
  func.func @main(%arg0: tensor<1x2xi32>) -> tensor<1x2x2xi32> {
  %0 = "tfl.custom"(%arg0) {custom_code = "stablehlo.broadcast_in_dim", custom_option = #tfl<const_bytes : "0x62726F6164636173745F64696D656E73696F6E73000201020119010101072C022401">} : (tensor<1x2xi32>) -> tensor<1x2x2xi32>
  func.return %0 : tensor<1x2x2xi32>
  }
}

// CHECK:      module {
// CHECK-NEXT:  func @main(%arg0: tensor<1x2xi32>) -> tensor<1x2x2xi32> {

    return %2 : tensor<1x3xf32>
  }

  func.func private @composite_dot_general_fn(%arg0: tensor<1x2xf32>, %arg1: tensor<2x3xf32>) -> tensor<1x3xf32> attributes {_from_xla_call_module} {
    %0 = stablehlo.dot_general %arg0, %arg1, contracting_dims = [1] x [0] : (tensor<1x2xf32>, tensor<2x3xf32>) -> tensor<1x3xf32>
    return %0 : tensor<1x3xf32>
  }
}

  %1 = "tf.MatMul"(%arg0, %0) {T = f32, device = "/device:CPU:0", transpose_a = false, transpose_b = false} : (tensor<3x1xf32>, tensor<1x3xf32>) -> tensor<3x3xf32>
  func.return %1 : tensor<3x3xf32>
}

// CHECK-LABEL: func @gpu_device
func.func @gpu_device(%arg0: tensor<3x1xf32>, %arg1: tensor<!tf_type.resource<tensor<1x3xf32>>>) -> tensor<3x3xf32> {

      _collective_manager_ids = [], device = ""
    } : (tensor<3x3xf32>, tensor<3x3xf32>) -> tensor<3x3xf32>
    %3 = "tf.PartitionedCall"(%2, %1) <{
      config = "", config_proto = "", executor_type = "", f = @some_other_func
    }> {
      _collective_manager_ids = [], device = ""
    } : (tensor<3x3xf32>, tensor<3x3xf32>) -> tensor<3x3xf32>
    return %3 : tensor<3x3xf32>
  }
  // CHECK: func.func @main