import thunder fails on torch >= 2.14.0: Function.apply has no __func__ (pytorch #190694 moved apply into C++)

[BwL1289]

What happens

On torch 2.14.0 (and later), import thunder raises at module import:

File ".../thunder/__init__.py", line 30, in <module>
    from thunder.core.jit_ext import thunder_general_jit
File ".../thunder/core/jit_ext.py", line 797, in <module>
    @register_general_jit_lookaside(torch.autograd.function.Function.apply.__func__)
AttributeError: 'builtin_function_or_method' object has no attribute '__func__'

Why

pytorch/pytorch#190694 (4fee7d8a70cfdd5fe88cf949cd9e95f67af9c369, merged 2026-07-23) moved the
autograd.Function.apply wrapper into C++. torch/autograd/function.py now reads

apply = _C._FunctionBase.__dict__["apply"]

so Function.apply is a builtin_function_or_method with no __func__. This is in the v2.14.0
tag (line 650); v2.13.0 still has def apply(cls, *args, **kwargs) (line 610).

Reproducer

pip install "torch==2.14.0" lightning-thunder
python -c "import thunder"

(On 2.14.0 you will hit the align_as failure below first unless the named-tensor references are
guarded; with them guarded, the traceback above is what you get.)

Why CI has not caught it (as of thunder main @ 422f45b)

• The latest lanes (11 jobs) never install: requirements/devel.txt asks for
  torchaudio==2.14.0, which does not exist, so pip fails before thunder is imported.
• The nightly lanes (7 jobs) fail earlier, in thunder/torch/default_torch_ops.py:361, with
  AttributeError: type object 'Tensor' has no attribute 'align_as' — Remove named tensor pytorch/pytorch#173895
  (00512145ec4) removed named tensors, and the module-level table evaluates
  torch.Tensor.{align_as,align_to,has_names,refine_names,rename} unguarded. Worth flagging on its
  own: that removal is in released torch 2.13.0, not just nightlies
  (test/test_namedtensor.py is present at v2.12.1 and gone at v2.13.0), so plain
  pip install torch==2.13.0 lightning-thunder; python -c "import thunder" already fails today.
• The oldest lane (3 jobs, torch==2.7.1) passes — the only green pytester jobs in that run.

We hit the apply failure only because we already carry a local hasattr guard for those five
named-tensor attributes; that guard is what gets the import past default_torch_ops.py to
jit_ext.py:797. Happy to send it as a separate PR if useful.

A re-key alone is not a fix

We measured two obvious one-line repairs — keying the lookaside on
torch.autograd.function.Function.apply, and on torch._C._FunctionBase.__dict__["apply"]. Both
make the import succeed, and with both the lookaside never fires for a user subclass:

MyFunc.apply in _general_jit_lookaside_map: False
MyFunc.apply.__self__:   <class '__main__.MyFunc'>
Function.apply.__self__: <class 'torch.autograd.function.Function'>

thunder.jit on a trivial torch.autograd.Function then dies in the opaque-function fallback:

thunder.core.interpreter.InterpreterError: Encountered exception KeyError: None while tracing ...
  thunder/core/jit_ext.py:1365 in general_jit_lookaside -> register_operator_for_opaque_function(fn)
  thunder/extend/__init__.py:416 in get_name -> mod = sys.modules[fn.__module__]
    KeyError: None

(A plain function under thunder.jit compiles and runs fine in the same process, so this is
specific to autograd.Function.)

Under the new torch, every subclass's .apply is a distinct bound builtin, so the interpreter
needs to recognise a builtin_function_or_method whose __self__ is an autograd.Function
subclass and route it to _general_jit_torch_autograd_function_apply_lookaside, rather than
matching a single shared function object. Separately, get_name could give a clearer error when
fn.__module__ is None.

Environment

lightning-thunder 0.2.7.dev0 (main @ 422f45b), CPython 3.14.7,
torch 2.14.0a0, aarch64. The jit_ext.py:797 line is also present at the latest release tag
(0.2.6), so this is not trunk-only.