Fix mypy errors attributed to pytorch_lightning.callbacks.quantization (#13782)

Co-authored-by: otaj <[email protected]>
Co-authored-by: Adrian Wälchli <[email protected]>

Author: 3 people

pyproject.toml

@@ -50,7 +50,6 @@ warn_no_return = "False"
 # mypy --no-error-summary 2>&1 | tr ':' ' ' | awk '{print $1}' | sort | uniq | sed 's/\.py//g; s|src/||g;  s|\/|\.|g' | xargs -I {} echo '"{}",'
 module = [
     "pytorch_lightning.callbacks.progress.rich_progress",
-    "pytorch_lightning.callbacks.quantization",
     "pytorch_lightning.core.datamodule",
     "pytorch_lightning.demos.boring_classes",
     "pytorch_lightning.demos.mnist_datamodule",

src/pytorch_lightning/callbacks/quantization.py

@@ -41,25 +41,28 @@
 
 
 def wrap_qat_forward_context(
-    quant_cb, model: "pl.LightningModule", func: Callable, trigger_condition: Optional[Union[Callable, int]] = None
+    quant_cb: "QuantizationAwareTraining",
+    model: "pl.LightningModule",
+    func: Callable,
+    trigger_condition: Optional[Union[Callable, int]] = None,
 ) -> Callable:
     """Decorator to wrap forward path as it is needed to quantize inputs and dequantize outputs for in/out
     compatibility Moreover this version has the (de)quantization conditional as it may not be needed for the
     training all the time."""
     # todo: consider using registering hook before/after forward
     @functools.wraps(func)
-    def wrapper(data) -> Any:
-        _is_func_true = isinstance(trigger_condition, Callable) and trigger_condition(model.trainer)
+    def wrapper(data: Any) -> Any:
+        _is_func_true = callable(trigger_condition) and trigger_condition(model.trainer)
         _is_count_true = isinstance(trigger_condition, int) and quant_cb._forward_calls < trigger_condition
         _quant_run = trigger_condition is None or _is_func_true or _is_count_true
         # apply custom trigger
         if _quant_run:
             quant_cb._forward_calls += 1
-            data = model.quant(data)
+            data = model.quant(data)  # type: ignore[operator]
         data = func(data)
         # apply custom trigger
         if _quant_run:
-            data = model.dequant(data)
+            data = model.dequant(data)  # type: ignore[operator]
         return data
 
     return wrapper
@@ -70,10 +73,10 @@ def wrap_quantize_forward_context(model: "pl.LightningModule", func: Callable) -
     compatibility."""
     # todo: consider using registering hook before/after forward
     @functools.wraps(func)
-    def wrapper(data) -> Any:
-        data = model.quant(data)
+    def wrapper(data: Any) -> Any:
+        data = model.quant(data)  # type: ignore[operator]
         data = func(data)
-        data = model.dequant(data)
+        data = model.dequant(data)  # type: ignore[operator]
         return data
 
     return wrapper
@@ -181,7 +184,9 @@ def __init__(
             )
         self._observer_type = observer_type
 
-        if collect_quantization is not None and not isinstance(collect_quantization, (int, Callable)):
+        if collect_quantization is not None and not (
+            isinstance(collect_quantization, int) or callable(collect_quantization)
+        ):
             raise MisconfigurationException(
                 f'Unsupported `collect_quantization` "{collect_quantization}", allowed are `int` or `Callable`.'
             )
@@ -200,8 +205,8 @@ def __init__(
         self._observer_disabled_stages = set(self.OBSERVER_STAGES) - observer_enabled_stages
 
         self._forward_calls = 0
-        self._fake_quant_to_initial_state_dict = {}
-        self._last_fake_quant_to_observer_enabled = {}
+        self._fake_quant_to_initial_state_dict: Dict[FakeQuantizeBase, Dict[str, Any]] = {}
+        self._last_fake_quant_to_observer_enabled: Dict[FakeQuantizeBase, Tensor] = {}
         self._module_prepared = False
 
     def _check_feasible_fuse(self, model: "pl.LightningModule") -> bool:
@@ -227,7 +232,7 @@ def _restore_last_observer_enabled(self) -> None:
         for fake_quant, observer_enabled in self._last_fake_quant_to_observer_enabled.items():
             fake_quant.observer_enabled.copy_(observer_enabled)
 
-    def _prepare_model(self, model: torch.nn.Module) -> None:
+    def _prepare_model(self, model: "pl.LightningModule") -> None:
         if self._module_prepared:
             return
         # QuantStub converts tensors from floating point to quantized
@@ -237,7 +242,7 @@ def _prepare_model(self, model: torch.nn.Module) -> None:
         # manually specify where tensors will be converted from quantized
         # to floating point in the quantized model
         self.__module_forward = model.forward
-        model.forward = wrap_qat_forward_context(
+        model.forward = wrap_qat_forward_context(  # type: ignore [assignment]
             quant_cb=self, model=model, func=model.forward, trigger_condition=self._collect_quantization
         )
 
@@ -247,7 +252,7 @@ def _prepare_model(self, model: torch.nn.Module) -> None:
             if self._observer_type == "histogram":
                 model.qconfig = torch.quantization.get_default_qconfig(self._qconfig)
             elif self._observer_type == "average":
-                extra_kwargs = {}
+                extra_kwargs: Dict[str, Optional[int]] = {}
                 if _TORCH_GREATER_EQUAL_1_12:
                     extra_kwargs["version"] = 0
                 # version=None corresponds to using FakeQuantize rather than
@@ -258,7 +263,7 @@ def _prepare_model(self, model: torch.nn.Module) -> None:
                 model.qconfig = torch.quantization.get_default_qat_qconfig(self._qconfig, **extra_kwargs)
 
         elif isinstance(self._qconfig, QConfig):
-            model.qconfig = self._qconfig
+            model.qconfig = self._qconfig  # type: ignore [assignment]
 
         if self._check_feasible_fuse(model):
             fuse_modules(model, self._modules_to_fuse, inplace=True)
@@ -273,12 +278,12 @@ def _prepare_model(self, model: torch.nn.Module) -> None:
         }
         self._module_prepared = True
 
-    def on_fit_start(self, trainer: "pl.Trainer", pl_module: "pl.LightningModule"):
+    def on_fit_start(self, trainer: "pl.Trainer", pl_module: "pl.LightningModule") -> None:
         self._prepare_model(pl_module)
 
     def on_fit_end(self, trainer: "pl.Trainer", pl_module: "pl.LightningModule") -> None:
         if not self._convert_on_fit_end:
-            pl_module.forward = self.__module_forward
+            pl_module.forward = self.__module_forward  # type: ignore [assignment]
             return
         pl_module.eval()
         # Convert the observed model to a quantized model. This does several things:
@@ -288,9 +293,12 @@ def on_fit_end(self, trainer: "pl.Trainer", pl_module: "pl.LightningModule") ->
         torch.quantization.convert(pl_module, inplace=True)
         # check we shall preserve wrapper
         if self._input_compatible:
-            pl_module.forward = wrap_quantize_forward_context(model=pl_module, func=self.__module_forward)
+            pl_module.forward = wrap_quantize_forward_context(  # type: ignore [assignment]
+                model=pl_module,
+                func=self.__module_forward,
+            )
         else:
-            pl_module.forward = self.__module_forward
+            pl_module.forward = self.__module_forward  # type: ignore [assignment]
 
     def on_train_start(self, trainer: "pl.Trainer", pl_module: "pl.LightningModule") -> None:
         if "train" in self._observer_disabled_stages:
@@ -336,7 +344,7 @@ def state_dict(self) -> Dict[str, Any]:
         keys = {"_qconfig", "_observer_type", "_collect_quantization", "_modules_to_fuse", "_input_compatible"}
         return {n: getattr(self, n) for n in keys}
 
-    def _load_before_model(self, model: torch.nn.Module, state_dict: Dict[str, Any]) -> None:
+    def _load_before_model(self, model: "pl.LightningModule", state_dict: Dict[str, Any]) -> None:
         """Special hook that gets called by the CheckpointConnector *before* the model gets loaded.
 
         This hook replaces the :meth:`on_load_checkpoint` and :meth:`load_state_dict` callback methods which get called

src/pytorch_lightning/trainer/connectors/checkpoint_connector.py

@@ -245,7 +245,7 @@ def _restore_quantization_callbacks(self) -> None:
             if state:
                 # The Quantization callbacks have a special method that must be called before restoring the weights
                 # of the model
-                callback._load_before_model(self.trainer.model, deepcopy(state))
+                callback._load_before_model(self.trainer.lightning_module, deepcopy(state))
 
     def restore_callbacks(self) -> None:
         """Restores all callbacks from the pre-loaded checkpoint."""