What precisely is required for isinstance checks to succeed against data-Protocols?

[randolf-scholz]

The following example fails at runtime:

from typing import Protocol, runtime_checkable  # python 3.13.2
from torch import nn  # v2.6.0

@runtime_checkable
class HasWeight(Protocol):
    weight: nn.Parameter

model = nn.Linear(3, 4)
assert hasattr(model, "weight")                   # ✅
assert isinstance(model.weight, nn.Parameter)     # ✅
assert "weight" in model.__static_attributes__    # ✅
assert "weight" in model.__annotations__          # ✅
assert isinstance(model, HasWeight)               # ❌

It seems the culprit is a combination of torch.nn.Module doing stuff with __getattr__/__setattr_ which makes inspect.getattr_static fail. However, what surprises me is that both __annotations__ and __static_attributes__ are seemingly ignored. Is this intended or just underspecified?

I am trying to replicate this from scratch, but even with some non-static indirection it still works, haven't figured out what exactly makes torch break the isinstance check here.

from typing import Any, Protocol, runtime_checkable


@runtime_checkable
class Proto(Protocol):
    param: float


class Impl:
    param: float

    def __init__(self) -> None:
        self.register_parameter("param", 1.0)

    def __setattr__(self, name: str, value: Any) -> None:
        if isinstance(value, float):
            self.register_parameter(name, value)
        else:
            super().__setattr__(name, value)

    def register_parameter(self, name: str, value: float) -> None:
        assert value >= 0.0, "Negative param not allowed"
        super().__setattr__(name, value)


model = Impl()
assert "param" in model.__annotations__            # ✅
assert "param" not in model.__static_attributes__  # ✅
assert isinstance(model, Proto)                    # ✅

[randolf-scholz]

So, the issue seems to be that torch does not write an entry into __dict__, but uses different dictionaries for buffers / parameters / sub-modules.

[JelleZijlstra]

The implementation of runtime-checkable protocols indeed uses getattr_static. I haven't tracked down why this doesn't work as expected with torch, but in general runtime-checkable protocols with non-method members are very tricky and fragile; I'd avoid using them.

[randolf-scholz]

Yes it is because torch uses specific dictionaries _parameters, _modules etc. instead of __dict__. I posted a reproduction in plain python in discourse. https://discuss.python.org/t/typing-spec-isinstance-checks-on-data-protocols-possible-changes/83369