Fix mypy errors (#54)

Summary:
- Added one minor type hint.
- Ignored errors about instantiating a class with an `abstractmethod` since this seems to be a shortcoming on mypy's side: it doesn't recognize that the `dataclass` decorator will take care of `__init__`.
- Removed a `super()` call in the `__post_init__` function of `AdaGradGraftingConfig`. I think this is not necessary and mypy complains about inheriting from an `abstractmethod`.

Pull Request resolved: #54

Reviewed By: anana10c

Differential Revision: D66241155

Pulled By: tsunghsienlee

fbshipit-source-id: 5cd918d7c02693d1fd941b99e1b5ae9e06532cad

distributed_shampoo/examples/trainer_utils.py

@@ -529,7 +529,7 @@ def instantiate_grafting_config(
             epsilon=grafting_epsilon,
         )
     elif grafting_type == GraftingType.SGD:
-        return SGDGraftingConfig(
+        return SGDGraftingConfig(  # type: ignore[abstract]
             beta2=grafting_beta2,
             epsilon=grafting_epsilon,
         )

distributed_shampoo/gpu_tests/shampoo_grafting_test.py

@@ -274,7 +274,7 @@ def test_sgd_grafting_on_quadratic(self) -> None:
                         start_preconditioning_step=math.inf,
                         use_nesterov=use_nesterov,
                         use_decoupled_weight_decay=False,
-                        grafting_config=SGDGraftingConfig(),
+                        grafting_config=SGDGraftingConfig(),  # type: ignore[abstract]
                     ),
                     device=device,
                 )

distributed_shampoo/shampoo_types.py

@@ -246,7 +246,6 @@ class AdaGradGraftingConfig(GraftingConfig):
     epsilon: float = 1e-10
 
     def __post_init__(self) -> None:
-        super().__init__()
         if not self.epsilon > 0.0:
             raise ValueError(f"Invalid epsilon value: {self.epsilon}. Must be > 0.0.")
 

distributed_shampoo/tests/distributed_shampoo_test.py

@@ -60,7 +60,7 @@ def test_invalid_grafting_config(self) -> None:
         ):
             DistributedShampoo(
                 self._model.parameters(),
-                grafting_config=SGDGraftingConfig(),
+                grafting_config=SGDGraftingConfig(),  # type: ignore[abstract]
             )
 
     def test_invalid_with_incorrect_hyperparameter_setting(self) -> None:

distributed_shampoo/utils/gpu_tests/shampoo_fully_shard_distributor_test.py

@@ -216,7 +216,7 @@ def _model_factory(
 
     @skip_if_lt_x_gpu(2)
     def test_fully_shard_shampoo_against_default_shampoo(self) -> None:
-        fully_shard_config = FullyShardShampooConfig()
+        fully_shard_config = FullyShardShampooConfig()  # type: ignore[abstract]
         ShampooFullyShardDistributorTest._test_two_configs(
             ShampooFullyShardDistributorTest._shampoo_optim_factory(
                 None,

distributed_shampoo/utils/tests/shampoo_preconditioner_list_test.py

@@ -282,7 +282,7 @@ def test_abstract_methods(self) -> None:
         # Basic setup for instantiating BaseShampooPreconditionerList.
         params = (torch.tensor([1.0, 2.0]),)
         block_list = (params[0],)
-        state = {params[0]: {}}
+        state: dict[Tensor, dict] = {params[0]: {}}
         block_info_list = (
             BlockInfo(
                 param=params[0],

tests/matrix_functions_test.py

@@ -302,7 +302,7 @@ def test_matrix_inverse_root_with_invalid_root_inv_config(self) -> None:
             matrix_inverse_root(
                 A=A,
                 root=root,
-                root_inv_config=InvalidRootInvConfig(),
+                root_inv_config=InvalidRootInvConfig(),  # type: ignore[abstract]
                 is_diagonal=False,
             )