🐛 Describe the bug

This turned up after https://github.com/pytorch/pytorch/pull/163654 was merged and included in PyTorch 2.10: Some tests use the now deprecated class in a JIT context which does not work. MWE:

from typing_extensions import deprecated
import torch

@deprecated("!")
class Foo:
    def __init__(self):
        pass

@torch.jit.script
def fn():
    f=Foo()

TypeError: module, class, method, function, traceback, frame, or code object was expected, got builtin_function_or_method

Cause: The __new__ method wrapper added by @deprecated causes it to be picked up by the JIT passes which then fails in 2 places:

1. https://github.com/pytorch/pytorch/blob/24cdfbdd4ce3d3b080e64e6a8b03a8ecda547bd0/torch/_jit_internal.py#L593 doesn't detect it as builtin anymore which makes inspect.getsource in inspect.getsource fail
2. It now appears in cls.__dict__ so it is not filtered anymore at https://github.com/pytorch/pytorch/blob/f605bd7dcc16ff97328db88b1005a9aecff2fed4/torch/jit/frontend.py#L263 which makes it fail in get_source_lines_and_files on calling inspect.getsourcefile(obj)

Test failure is reproducible by running python test/test_jit_autocast.py TestAutocast.test_callees on a CUDA build with Python 3.13, the snipped above is a reduced sample. This suggests there is no CI using Python 3.13 or this would have been detected earlier.

Partial fix is using inspect.unwrap before further processing which would yield the underlying method, e.g. object.__new__ in this case. But I'd say that the wrapped method is virtually never the one meant to be inspected: A signature of *args, **kwargs as usually used by wrappers is too generic to be helpful for JIT

Versions

Test failures introduced in PyTorch 2.10, still present in 2.12-rc9

The actual non-support was already present in 2.5, haven't tested even older versions.

cc @EikanWang @jgong5 @wenzhe-nrv @sanchitintel