return %2 : tensor<1x3xf32>
  }
  // CHECK-LABEL: func.func private @composite_dot_general_fn_1
  // CHECK-SAME: -> tensor<1x3xf32>
  func.func private @composite_dot_general_fn_1(%arg0: tensor<1x1024xf32>, %arg1: tensor<1024x3xf32>) -> tensor<1x3xf32> attributes {_from_xla_call_module} {

  // MLIR @main function corresponds to the TF function "main_original".
  OwningOpRef<ModuleOp> module_op = ParseModuleOpString(R"mlir(
    func.func private @main(%arg: tensor<1x2xf32>) -> (tensor<1x2xf32>) attributes {tf._original_func_name = "main_original"} {
      return %arg : tensor<1x2xf32>
    }
  )mlir");
  ASSERT_TRUE(module_op);

  absl::flat_hash_map<FunctionName, FunctionAlias> function_aliases;

module {
  // CHECK-LABEL: func @main
  // CHECK-SAME: %[[ARG_0:.*]]: tensor<10xf32>, %[[ARG_1:.*]]: tensor<10xf32>, %[[ARG_2:.*]]: tensor<6xi32>
  func.func @main(
      %arg0: tensor<10xf32>, %arg1: tensor<10xf32>, %arg2: tensor<6xi32>
    ) -> (tensor<10xf32>, tensor<6xi32>) {
    // CHECK: %[[CAST_0:.*]] = "tf.Cast"(%[[ARG_0]]) <{Truncate = false}> : (tensor<10xf32>) -> tensor<10xbf16>

  %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

  %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

  %5 = shape.shape_of %4 : tensor<?x10xf32> -> tensor<2xindex>
  %6 = stablehlo.dynamic_broadcast_in_dim %1, %5, dims = [1] : (tensor<10xf32>, tensor<2xindex>) -> tensor<?x10xf32>
  %7 = stablehlo.add %4, %6 : tensor<?x10xf32>
  %8 = stablehlo.clamp %2, %7, %3 : (tensor<f32>, tensor<?x10xf32>, tensor<f32>) -> tensor<?x10xf32>
  func.return %8: tensor<?x10xf32>
}

          %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 = "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>