Feature: Supressing unbound typevar warning

[mirceamironenco]

Is there a way to suppress the warning from an unbound typevar?

from collections.abc import Mapping
from dataclasses import dataclass
from typing import Annotated, Callable, Generic

import tyro
from typing_extensions import TypedDict, TypeVar

MapItemT = TypeVar("MapItemT", bound=Mapping)


@dataclass(kw_only=True, frozen=True)
class HFPromptDatasetConfig(Generic[MapItemT]):
    path: str
    prompt_keymap: Annotated[str | Callable[[MapItemT], str], tyro.conf.Suppress]
    answer_keymap: Annotated[str | Callable[[MapItemT], str], tyro.conf.Suppress]

CountdownItem = TypedDict[{"target": int, "nums": list[int]}]

def qwen_prompt(item: CountdownItem) -> str:
    return "foo"


@dataclass
class Config:
    train_data: HFPromptDatasetConfig = HFPromptDatasetConfig(
        path="Jiayi-Pan/Countdown-Tasks-3to4",
        prompt_keymap=qwen_prompt,
        answer_keymap="target",
    )


def main(config: Config) -> None:
    print(config)


if __name__ == "__main__":
    config = tyro.cli(Config)

    main(config)

Note that I've tried to Supress the fields that use the typevar but it doesn't help.
tyro will warn UserWarning: Could not resolve type parameter ~MapItemT. Type parameter resolution is not always possible in @staticmethod or @classmethod., however even though the generic parameter is not set (as in writing HFPromptDatasetConfig[CountdownItem]) the type checker understands that CountdownItem is bound the MapItemT, so I see a mismatch here. For example:

from collections.abc import Mapping
from dataclasses import dataclass
from typing import Annotated, Callable, Generic

import tyro
from typing_extensions import TypedDict, TypeVar

MapItemT = TypeVar("MapItemT", bound=Mapping)


@dataclass(kw_only=True, frozen=True)
class HFPromptDatasetConfig(Generic[MapItemT]):
    path: str
    prompt_keymap: Annotated[str | Callable[[MapItemT], str], tyro.conf.Suppress]
    answer_keymap: Annotated[str | Callable[[MapItemT], str], tyro.conf.Suppress]

CountdownItem = TypedDict[{"target": int, "nums": list[int]}]
SomeOtherItem = TypedDict[{"foo": int, "bar": list[int]}]


def qwen_prompt(item: CountdownItem) -> str:
    return "foo"

def foo_answer(item: SomeOtherItem) -> str: ...


@dataclass
class Config:
    train_data: HFPromptDatasetConfig = HFPromptDatasetConfig(
        path="Jiayi-Pan/Countdown-Tasks-3to4",
        prompt_keymap=qwen_prompt,
        answer_keymap=foo_answer,
    )


def main(config: Config) -> None:
    print(config)


if __name__ == "__main__":
    config = tyro.cli(Config)

    main(config)

Pyright will issue a correct warning here for answer_keymap=foo_answer:

Argument of type "(item: SomeOtherItem) -> str" cannot be assigned to parameter "answer_keymap" of type "str | ((MapItemT@HFPromptDatasetConfig) -> str)" in function "__init__"
  Type "(item: SomeOtherItem) -> str" is not assignable to type "str | ((CountdownItem) -> str)"
    "FunctionType" is not assignable to "str"
    Type "(item: SomeOtherItem) -> str" is not assignable to type "(CountdownItem) -> str"
      Parameter 1: type "CountdownItem" is incompatible with type "SomeOtherItem"
        "foo" is missing from "<TypedDict>"
        "bar" is missing from "<TypedDict>"

Even if I use a TypeVar with a default the warning is still issued:

from collections.abc import Mapping
from dataclasses import dataclass
from typing import Annotated, Callable, Generic

import tyro
from typing_extensions import TypeVar

MapItemT = TypeVar("MapItemT", bound=Mapping, default=Mapping)


@dataclass(kw_only=True, frozen=True)
class HFPromptDatasetConfig(Generic[MapItemT]):
    path: str
    prompt_keymap: Annotated[str | Callable[[MapItemT], str], tyro.conf.Suppress]
    answer_keymap: Annotated[str | Callable[[MapItemT], str], tyro.conf.Suppress]


@dataclass
class Config:
    train_data: HFPromptDatasetConfig = HFPromptDatasetConfig(
        path="Jiayi-Pan/Countdown-Tasks-3to4",
        prompt_keymap="nums",
        answer_keymap="target",
    )


def main(config: Config) -> None:
    print(config)


if __name__ == "__main__":
    config = tyro.cli(Config)

    main(config)

If there is no way to Suppress I would be happy at least if the warnings.warn(...) used a category= and tyro had a specific warning category here that I can then ignore myself, since other warnings might be relevant.

[brentyi]

Interesting, thanks @mirceamironenco!

I think we can:

1. Support default= in the TypeVar; the warning shouldn't be raised when this is set
2. In general: define a custom warning category for warnings in tyro

Would these solve your problem?

[mirceamironenco]

Yes, that would work for me. Thanks for taking a look at this!