Python/ruff
1
0
Fork 0
mirror of https://github.com/astral-sh/ruff synced 2026-01-06 22:24:05 -05:00
Code Issues Packages Projects Releases Wiki Activity

[ty] Treat Callable dunder members as bound method descriptors (#20860)

Browse Source 
## Summary

Dunder methods (at least the ones defined in the standard library)
always take an instance of the class as the first parameter. So it seems
reasonable to generally treat them as bound method descriptors if they
are defined via a `Callable` type.

This removes just a few false positives from the ecosystem, but solves
three user-reported issues:

closes https://github.com/astral-sh/ty/issues/908
closes https://github.com/astral-sh/ty/issues/1143
closes https://github.com/astral-sh/ty/issues/1209

In addition to the change here, I also considered [making `ClassVar`s
bound method descriptors](https://github.com/astral-sh/ruff/pull/20861).
However, there was zero ecosystem impact. So I think we can also close
https://github.com/astral-sh/ty/issues/491 with this PR.

closes https://github.com/astral-sh/ty/issues/491

## Test Plan

Added regression test
This commit is contained in:
David Peter
2025-10-14 14:27:52 +02:00
committed by GitHub
parent ac2c530377
commit 6341bb7403
5 changed files with 82 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

34
crates/ty_python_semantic/resources/mdtest/call/callables_as_descriptors.md
View File

@@ -204,3 +204,37 @@ class Calculator:
reveal_type(Calculator().square_then_round(3.14)) # revealed: Unknown | int
```
## Use case: Treating dunder methods as bound-method descriptors
pytorch defines a `__pow__` dunder attribute on [`TensorBase`] in a similar way to the following
example. We generally treat dunder attributes as bound-method descriptors since they all take a
`self` argument. This allows us to type-check the following code correctly:
```py
from typing import Callable
def pow_impl(tensor: Tensor, exponent: int) -> Tensor:
raise NotImplementedError
class Tensor:
__pow__: Callable[[Tensor, int], Tensor] = pow_impl
Tensor() ** 2
```
The following example is also taken from a real world project. Here, the `__lt__` dunder attribute
is not declared. The attribute type is therefore inferred as `Unknown | Callable[…]`, but we still
treat it as a bound-method descriptor:
```py
def make_comparison_operator(name: str) -> Callable[[Matrix, Matrix], bool]:
raise NotImplementedError
class Matrix:
__lt__ = make_comparison_operator("lt")
Matrix() < Matrix()
```
[`tensorbase`]: https://github.com/pytorch/pytorch/blob/f3913ea641d871f04fa2b6588a77f63efeeb9f10/torch/_tensor.py#L1084-L1092