no-member false positive with super() with huggingface's transformers.Trainer

[Hnasar]

Bug description

Pylint seems to improperly parse transformers.Trainer so it issues a false positive no-member for this code:

# test_module.py
from transformers import Trainer

class B(Trainer):
    def evaluation_loop(self):
        super().evaluation_loop()
        print('bye')

but the method definitely exists.

I thought it was maybe astroid failing to parse the Trainer class which has inline function imports

    from .trainer_pt_utils import _get_learning_rate, log_metrics, metrics_format, save_metrics, save_state

because I tried putting a breakpoint in pylint/checkers/typecheck.py before the _emit_no_member and noticing that owner.super_mro()[0].locals was missing nearly all of the methods,

but then I manually tried parsing the transformers/trainer.py then the method is defined.

import astroid
from pathlib import Path
print(list(astroid.parse(Path("venv/lib/python3.11/site-packages/transformers/trainer.py").read_text()).get_children())[-1].locals)

So there seems to be some other bug with how pylint is analyzing this code.

Configuration

No response

Command used

pylint test_module.py

Pylint output

************* Module test_module
test_module.py:5:8: E1101: Super of 'B' has no 'evaluation_loop' member (no-member)

Expected behavior

No error.

Pylint version

pylint 4.0.0-dev0
astroid 3.3.2
Python 3.11.10 (main, Sep  9 2024, 05:30:15) [GCC 11.2.0]

OS / Environment

Running from commit
0972ba5

but it's also reproducible with pylint 3.2.6

Additional dependencies

transformers==4.45.0

[Hnasar]

In further investigation, it was pointed out to me that there's an alternate Trainer class
https://github.com/huggingface/transformers/blob/main/src/transformers/utils/dummy_pt_objects.py#L10029

# transformers/__init__.py
    try:
        if not is_torch_available():
            raise OptionalDependencyNotAvailable()
    except OptionalDependencyNotAvailable:
        from .utils.dummy_pt_objects import *   # alternate Trainer class
    else:
        …
        from .trainer import Trainer

# transformers/utils/dummy_pt_objects.py
class Trainer(metaclass=DummyObject):
    _backends = ["torch"]

    def __init__(self, *args, **kwargs):
        requires_backends(self, ["torch"])

So the workaround is to explicitly import from transformers.trainer import Trainer but still it would be nice if pylint handled this case better — maybe it could always assume is_torch_available is True