Incorrect detection of generics with nested classes

[M0dEx]

Initial Checks

• I confirm that I'm using Pydantic V2

Description

When nesting generic classes, one of which is a subclass of BaseModel (as shown in the example), Pydantic incorrectly raises a TypeError with the explanation that one of the classes is not generic:

Traceback (most recent call last):
  File "/Users/***/Library/Application Support/JetBrains/Toolbox/apps/PyCharm-P/ch-0/232.8660.197/PyCharm.app/Contents/plugins/python/helpers/pydev/pydevconsole.py", line 364, in runcode
    coro = func()
           ^^^^^^
  File "<input>", line 12, in <module>
  File "/Users/***/Library/Caches/pypoetry/virtualenvs/***/lib/python3.11/site-packages/pydantic/main.py", line 607, in __class_getitem__
    raise TypeError(f'{cls} is not a generic class')
TypeError: <class '__main__.BrokenB'> is not a generic class

According to MyPy docs, the example should be valid (and without inheriting BaseModel works just fine.)

There is a workaround - by specifying the Generic keyword with the same TypeVar when inheriting the base class (A in the example code below), the check inside Pydantic succeeds.

Example Code

from typing import Generic, TypeVar

from pydantic import BaseModel

T = TypeVar("T")


class A(BaseModel, Generic[T]):
    a: T
    

class BrokenB(A[T]):
    b: T


class BrokenC(BrokenB[int]):  # FAILS
    pass


class FixedB(A[T], Generic[T]):
    b: T


class FixedC(FixedB[int]):  # WORKS
    pass

Python, Pydantic & OS Version

pydantic version: 2.1.1
pydantic-core version: 2.4.0
pydantic-core build: profile=release pgo=false mimalloc=true
install path: /Users/***/Library/Caches/pypoetry/virtualenvs/***-py3.11/lib/python3.11/site-packages/pydantic
python version: 3.11.2 (main, Mar 21 2023, 13:41:33) [Clang 14.0.0 (clang-1400.0.29.202)]
platform: macOS-13.5-x86_64-i386-64bit
optional deps. installed: ['typing-extensions']

Selected Assignee: @dmontagu

[dmontagu]

Up to this point, the fact that this didn't work was more of an implementation detail than anything intentional, but I think I found a way to make it work. At least, I found a tweak in #7119 that makes this work properly and keeps all existing tests passing. If this introduces any new issues, I'll address them when reported.