// were:
      //   tensor<!tf_type.variant<tensor<?x8xf32>>>
      // and:
      //   tensor<!tf_type.variant<tensor<10x8xf32>>>
      // we'll try here to refine tensor<?x8xf32> with tensor<10x8xf32>.
      auto refined_subtype = mlir::cast<TensorType>(
          TypeMeet(lhs_element_type_with_subtype.GetSubtypes().front(),

      tf_device.return %5 : tensor<512xi32>
    }) {device = "/job:localhost/replica:0/task:0/device:CPU:0"} : () -> (tensor<512xi32>, tensor<1024xi32>, tensor<1024xi32>, tensor<1024xf32>)
    // CHECK: %[[COMPILE_OUTPUT:[0-9]*]]:4 = "tf_device.launch"
    // CHECK: "tf._TPUCompileMlir"()
    // CHECK: is_bounded_dynamic_dim: true

mlir_module = '''python
func @main(%arg0 : tensor<10xf32>, %arg1 : tensor<10xf32>) -> tensor<10x10xf32> {
   %add = "magic.op"(%arg0, %arg1) : (tensor<10xf32>, tensor<10xf32>) -> tensor<10x10xf32>
   return %ret : tensor<10x10xf32>
}
'''

@tf.function
def foo(x, y):
  return mlir_passthrough_op([x, y], mlir_module, Toutputs=[tf.float32])