Fix attribute inheritance regression (introduced in 0.9.9) (#240)

* Fix attribute inheritance regression

* Add test

* Workaround for Python 3.8 __annotations__ behavior

* Remove unnecessary filter

* Additional test

* ruff

src/tyro/_resolver.py

@@ -673,7 +673,13 @@ def get_hints_for_bound_method(cls) -> Dict[str, Any]:
             for x, t in _get_type_hints_backported_syntax(
                 obj, include_extras=include_extras
             ).items()
-            if x in obj.__annotations__
+            # Only include type hints that are explicitly defined in this class.
+            #
+            # Why `cls.__dict__.__annotations__` instead of `cls.__annotations__`?
+            # Because in Python 3.8 and earlier, `cls.__annotations__`
+            # recursively merges parent class annotations.
+            # See this issue: https://github.com/python/cpython/issues/99535
+            if x in obj.__dict__.get("__annotations__", {})
         }
 
     # We need to recurse into base classes in order to correctly resolve superclass parameters.
@@ -689,6 +695,11 @@ def get_hints_for_bound_method(cls) -> Dict[str, Any]:
                 {
                     x: TypeParamResolver.concretize_type_params(t)
                     for x, t in base_hints.items()
+                    # Include type hints that are not assigned earlier in the MRO.
+                    #
+                    # This needs to be recursive (include parents of parents),
+                    # so we shouldn't filter by local __annotations__.
+                    if x not in out
                 }
             )
 

tests/test_conf.py

@@ -1794,3 +1794,94 @@ def main(
     with pytest.raises(SystemExit):
         # ConfigB has a required argument.
         assert tyro.cli(main, args=["x:config-b"])
+
+
+_dataset_map = {
+    "alpaca": "tatsu-lab/alpaca",
+    "alpaca_clean": "yahma/alpaca-cleaned",
+    "alpaca_gpt4": "vicgalle/alpaca-gpt4",
+}
+_inv_dataset_map = {value: key for key, value in _dataset_map.items()}
+_datasets = list(_dataset_map.keys())
+
+HFDataset = Annotated[
+    str,
+    tyro.constructors.PrimitiveConstructorSpec(
+        nargs=1,
+        metavar="{" + ",".join(_datasets) + "}",
+        instance_from_str=lambda args: _dataset_map[args[0]],
+        is_instance=lambda instance: isinstance(instance, str)
+        and instance in _inv_dataset_map,
+        str_from_instance=lambda instance: [_inv_dataset_map[instance]],
+        choices=tuple(_datasets),
+    ),
+    tyro.conf.arg(
+        help_behavior_hint=lambda df: f"(default: {df}, run datasets.py for full options)"
+    ),
+]
+
+
+def test_annotated_attribute_inheritance() -> None:
+    """From @mirceamironenco.
+
+    https://github.com/brentyi/tyro/issues/239"""
+
+    @dataclasses.dataclass(frozen=True)
+    class TrainConfig:
+        dataset: str = "vicgalle/alpaca-gpt4"
+
+    @dataclasses.dataclass(frozen=True)
+    class CLITrainerConfig(TrainConfig):
+        dataset: HFDataset = "vicgalle/alpaca-gpt4"
+
+    assert "{alpaca,alpaca_clean,alpaca_gpt4}" in get_helptext_with_checks(
+        CLITrainerConfig
+    )
+    assert (
+        "default: alpaca_gpt4, run datasets.py for full options"
+        in get_helptext_with_checks(CLITrainerConfig)
+    )
+
+
+@dataclasses.dataclass(frozen=True)
+class OptimizerConfig:
+    lr: float = 1e-1
+
+
+@dataclasses.dataclass(frozen=True)
+class AdamConfig(OptimizerConfig):
+    adam_foo: float = 1.0
+
+
+@dataclasses.dataclass(frozen=True)
+class SGDConfig(OptimizerConfig):
+    sgd_foo: float = 1.0
+
+
+@dataclasses.dataclass
+class TrainConfig:
+    optimizer: OptimizerConfig = AdamConfig()
+
+
+def _dummy_constructor() -> Type[OptimizerConfig]:
+    return Union[AdamConfig, SGDConfig]  # type: ignore
+
+
+CLIOptimizerConfig = Annotated[
+    OptimizerConfig,
+    tyro.conf.arg(constructor_factory=_dummy_constructor),
+]
+
+
+def test_attribute_inheritance_2() -> None:
+    """From @mirceamironenco.
+
+    https://github.com/brentyi/tyro/issues/239"""
+
+    @dataclasses.dataclass
+    class CLITrainerConfig(TrainConfig):
+        optimizer: CLIOptimizerConfig = SGDConfig()
+
+    assert "[{optimizer:adam-config,optimizer:sgd-config}]" in get_helptext_with_checks(
+        CLITrainerConfig
+    )

tests/test_py311_generated/test_conf_generated.py

@@ -1800,3 +1800,94 @@ def main(
     with pytest.raises(SystemExit):
         # ConfigB has a required argument.
         assert tyro.cli(main, args=["x:config-b"])
+
+
+_dataset_map = {
+    "alpaca": "tatsu-lab/alpaca",
+    "alpaca_clean": "yahma/alpaca-cleaned",
+    "alpaca_gpt4": "vicgalle/alpaca-gpt4",
+}
+_inv_dataset_map = {value: key for key, value in _dataset_map.items()}
+_datasets = list(_dataset_map.keys())
+
+HFDataset = Annotated[
+    str,
+    tyro.constructors.PrimitiveConstructorSpec(
+        nargs=1,
+        metavar="{" + ",".join(_datasets) + "}",
+        instance_from_str=lambda args: _dataset_map[args[0]],
+        is_instance=lambda instance: isinstance(instance, str)
+        and instance in _inv_dataset_map,
+        str_from_instance=lambda instance: [_inv_dataset_map[instance]],
+        choices=tuple(_datasets),
+    ),
+    tyro.conf.arg(
+        help_behavior_hint=lambda df: f"(default: {df}, run datasets.py for full options)"
+    ),
+]
+
+
+def test_annotated_attribute_inheritance() -> None:
+    """From @mirceamironenco.
+
+    https://github.com/brentyi/tyro/issues/239"""
+
+    @dataclasses.dataclass(frozen=True)
+    class TrainConfig:
+        dataset: str = "vicgalle/alpaca-gpt4"
+
+    @dataclasses.dataclass(frozen=True)
+    class CLITrainerConfig(TrainConfig):
+        dataset: HFDataset = "vicgalle/alpaca-gpt4"
+
+    assert "{alpaca,alpaca_clean,alpaca_gpt4}" in get_helptext_with_checks(
+        CLITrainerConfig
+    )
+    assert (
+        "default: alpaca_gpt4, run datasets.py for full options"
+        in get_helptext_with_checks(CLITrainerConfig)
+    )
+
+
+@dataclasses.dataclass(frozen=True)
+class OptimizerConfig:
+    lr: float = 1e-1
+
+
+@dataclasses.dataclass(frozen=True)
+class AdamConfig(OptimizerConfig):
+    adam_foo: float = 1.0
+
+
+@dataclasses.dataclass(frozen=True)
+class SGDConfig(OptimizerConfig):
+    sgd_foo: float = 1.0
+
+
+@dataclasses.dataclass
+class TrainConfig:
+    optimizer: OptimizerConfig = AdamConfig()
+
+
+def _dummy_constructor() -> Type[OptimizerConfig]:
+    return AdamConfig | SGDConfig  # type: ignore
+
+
+CLIOptimizerConfig = Annotated[
+    OptimizerConfig,
+    tyro.conf.arg(constructor_factory=_dummy_constructor),
+]
+
+
+def test_attribute_inheritance_2() -> None:
+    """From @mirceamironenco.
+
+    https://github.com/brentyi/tyro/issues/239"""
+
+    @dataclasses.dataclass
+    class CLITrainerConfig(TrainConfig):
+        optimizer: CLIOptimizerConfig = SGDConfig()
+
+    assert "[{optimizer:adam-config,optimizer:sgd-config}]" in get_helptext_with_checks(
+        CLITrainerConfig
+    )