AttributeError: 'builtin_function_or_method' object has no attribute '__python_implementation__'

[weifengpy]

🐛 Describe the bug

FakeTensorMode throw runtime error in latest build 7281e0b
TorchTitan use FakeTensorMode to estimate memory usage and that's why I found this

repro

from torch._subclasses.fake_tensor import FakeTensorMode
with FakeTensorMode():
    pass

Traceback (most recent call last):
  File "/data/users/weif/torchtitan/test_fake_tensor_mode.py", line 3, in <module>
    with FakeTensorMode():
  File "/data/users/weif/pytorch/torch/_subclasses/fake_tensor.py", line 1150, in __init__
    import torch._dynamo.config
  File "/data/users/weif/pytorch/torch/_dynamo/__init__.py", line 42, in <module>
    from .polyfills import loader as _  # usort: skip # noqa: F401
  File "/data/users/weif/pytorch/torch/_dynamo/polyfills/loader.py", line 24, in <module>
    POLYFILLED_MODULES: Tuple["ModuleType", ...] = tuple(
  File "/data/users/weif/pytorch/torch/_dynamo/polyfills/loader.py", line 25, in <genexpr>
    importlib.import_module(f".{submodule}", package=polyfills.__name__)
  File "/home/weif/.conda/envs/pytorch-3.10/lib/python3.10/importlib/__init__.py", line 126, in import_module
    return _bootstrap._gcd_import(name[level:], package, level)
  File "/data/users/weif/pytorch/torch/_dynamo/polyfills/pytree.py", line 52, in <module>
    __func.__python_implementation__
AttributeError: 'builtin_function_or_method' object has no attribute '__python_implementation__'

Versions

7281e0b

cc @chauhang @penguinwu

[weifengpy]

cc @bdhirsh in case you have context

[weifengpy]

this dynamo PR seems to be the root cause: #137397

thus adding PT2 oncall here

[jansel]

I am unable to repro this:

>>> from torch._subclasses.fake_tensor import FakeTensorMode
>>> with FakeTensorMode():
...     pass
... 
>>> 

running on 90052a8 (from earlier today)

[XuehaiPan]

Traceback (most recent call last):
  File "/data/users/weif/torchtitan/test_fake_tensor_mode.py", line 3, in <module>
    with FakeTensorMode():
  File "/data/users/weif/pytorch/torch/_subclasses/fake_tensor.py", line 1150, in __init__
    import torch._dynamo.config
  File "/data/users/weif/pytorch/torch/_dynamo/__init__.py", line 42, in <module>
    from .polyfills import loader as _  # usort: skip # noqa: F401
  File "/data/users/weif/pytorch/torch/_dynamo/polyfills/loader.py", line 24, in <module>
    POLYFILLED_MODULES: Tuple["ModuleType", ...] = tuple(
  File "/data/users/weif/pytorch/torch/_dynamo/polyfills/loader.py", line 25, in <genexpr>
    importlib.import_module(f".{submodule}", package=polyfills.__name__)
  File "/home/weif/.conda/envs/pytorch-3.10/lib/python3.10/importlib/__init__.py", line 126, in import_module
    return _bootstrap._gcd_import(name[level:], package, level)
  File "/data/users/weif/pytorch/torch/_dynamo/polyfills/pytree.py", line 52, in <module>
    __func.__python_implementation__
AttributeError: 'builtin_function_or_method' object has no attribute '__python_implementation__'

You need to upgrade optree:

python3 -m pip install 'optree>=0.13.0'

pytorch/requirements.txt

Line 20 in 2d9b081

optree>=0.13.0

Maybe we need to add a version guard in the polyfill.

[jansel]

@XuehaiPan can we make the error message better here or remove the requirement?

[weifengpy]

cc @XilunWu regarding python3 -m pip install 'optree>=0.13.0'

[E-Loba]

Hello!
Sorry to bring up an old issue. But I ran into the same problem with
torch==2.6
torchvision==0.21.0
optree==0.14.1
I was able to solve it by DOWNgrading to optree==0.13.0

[XuehaiPan]

Hi @E-Loba, have you tried to reinstall the package?

pip install -I optree