%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

# CHECK-NEXT: [[R0:%.*]] = "tf.ReadVariableOp"([[ARG1]]) {{{.*}}} : (tensor<!tf_type.resource<tensor<1x3xf32>>>) -> tensor<1x3xf32>
# CHECK-NEXT: [[R1:%.*]] = "tf.MatMul"([[ARG0]], [[R0]]) <{{{.*}}}> {{{.*}}} : (tensor<3x1xf32>, tensor<1x3xf32>) -> tensor<3x3xf32>
# CHECK-NEXT: return [[R1]] : tensor<3x3xf32>


def Test():

  x = tf.constant([[1.0], [1.0], [1.0]])
  y = tf.compat.v1.get_variable(
      name='y',

  %1 = "tf.ReadVariableOp"(%0) {device = "/device:CPU:0"} : (tensor<!tf_type.resource<tensor<1x3xf32>>>) -> tensor<1x3xf32>
  %2 = "tf.AddV2"(%arg0, %1) {device = "/device:CPU:0"} : (tensor<1x3xf32>, tensor<1x3xf32>) -> tensor<1x3xf32>
  %3 = "tf.Identity"(%2) {device = "/device:CPU:0"} : (tensor<1x3xf32>) -> tensor<1x3xf32>
  func.return %3 : tensor<1x3xf32>
}

// CHECK-LABEL: func @main

func.func @main(tensor<4x5xbf16>) -> tensor<4x5xbf16> {
^bb0(%arg0: tensor<4x5xbf16>):
  // CHECK-LABEL: @main
  // CHECK:  (tensor<4x5xbf16>) -> tensor<4x5xf32>
  // CHECK-NEXT:  (tensor<4x5xf32>) -> tensor<4x5xbf16>
  %0 = "tfl.cast" (%arg0) : (tensor<4x5xbf16>) -> tensor<4x5xf32> loc("cast1")
  %1 = "tfl.cast" (%0) : (tensor<4x5xf32>) -> tensor<4x5xbf16> loc("cast2")
  func.return %1 : tensor<4x5xbf16>

  %1 = "tfl.quantize"(%0) {qtype = tensor<2x1x!quant.uniform<i8:f32, 0.024986599940879671:92>>} : (tensor<2x1xf32>) -> tensor<2x1x!quant.uniform<i8:f32, 0.024986599940879671:92>>
  %2 = "tfl.dequantize"(%1) : (tensor<2x1x!quant.uniform<i8:f32, 0.024986599940879671:92>>) -> tensor<2x1xf32>

          %outputs_5, %control_6 = tf_executor.island(%control_4) wraps "tf._XlaHostComputeMlir"() {host_mlir_module = "module {\0A func.func @host_func() -> tensor<1x2xf32> {\0A %0 = \22tf.Const\22() {value = dense<0.1> : tensor<1x2xf32>} : () -> tensor<1x2xf32> \0A return %0 : tensor<1x2xf32>}}", manual_sharding = true, recv_key = "host_compute_channel_1_retvals", send_key = "host_c...

  %1 = "tf.Transpose"(%0, %cst_0): (tensor<1x2xf32>, tensor<2xi32>) -> tensor<2x1xf32>
  %2 = "tf.Transpose"(%0, %cst_0): (tensor<1x2xf32>, tensor<2xi32>) -> tensor<2x1xf32>
  func.return %1, %2 : tensor<2x1xf32>, tensor<2x1xf32>

// CHECK:  %cst = arith.constant

  %1 = "tf.Transpose"(%0, %cst_0): (tensor<1x2xf32>, tensor<2xi32>) -> tensor<2x1xf32>
  %2 = "tf.Transpose"(%0, %cst_0): (tensor<1x2xf32>, tensor<2xi32>) -> tensor<2x1xf32>
  func.return %1, %2 : tensor<2x1xf32>, tensor<2x1xf32>

// CHECK:  %cst = arith.constant

      func.func @main(%arg: tensor<1x2xf32> {tf_saved_model.index_path = ["input_tensor:0"]}) -> (tensor<1x2xf32> {tf_saved_model.index_path = ["output_tensor:0"]}) attributes {tf.entry_function = {inputs = "input_tensor:0", outputs = "output_tensor:0"}, tf_saved_model.exported_names = ["main"]} {
        %0 = tf_executor.graph {
          tf_executor.fetch %arg : tensor<1x2xf32>
        }
        return %0 : tensor<1x2xf32>
      }