fix defered annotations

python/cog/predictor.py

@@ -291,10 +291,7 @@ def get_input_create_model_kwargs(
     order = 0
 
     for name, parameter in signature.parameters.items():
-        if name not in input_types:
-            raise TypeError(f"No input type provided for parameter `{name}`.")
-
-        InputType = input_types[name]  # pylint: disable=invalid-name
+        InputType = input_types.get(name, parameter.annotation)
 
         validate_input_type(InputType, name)
 
@@ -360,7 +357,10 @@ class Input(BaseModel):
     predict = get_predict(predictor)
     signature = inspect.signature(predict)
 
-    input_types = get_type_hints(predict)
+    try:
+        input_types = get_type_hints(predict)
+    except TypeError:
+        input_types = {}
     if "return" in input_types:
         del input_types["return"]
 
@@ -374,16 +374,25 @@ class Input(BaseModel):
     )  # type: ignore
 
 
+def get_return_annotation(fn: Callable[..., Any]) -> Optional[Type[Any]]:
+    try:
+        return get_type_hints(fn).get("return", None)
+    except TypeError:
+        return_annotation = inspect.signature(fn).return_annotation
+        if return_annotation is inspect.Signature.empty:
+            return None
+        return return_annotation
+
+
 def get_output_type(predictor: BasePredictor) -> Type[BaseModel]:
     """
     Creates a Pydantic Output model from the return type annotation of a Predictor's predict() method.
     """
 
     predict = get_predict(predictor)
+    maybe_output_type = get_return_annotation(predict)
 
-    input_types = get_type_hints(predict)
-
-    if "return" not in input_types:
+    if maybe_output_type is None:
         raise TypeError(
             """You must set an output type. If your model can return multiple output types, you can explicitly set `Any` as the output type.
 
@@ -398,8 +407,8 @@ def predict(
         ...
 """
         )
-    OutputType = input_types.pop("return")  # pylint: disable=invalid-name
-
+    # we need the indirection to narrow the type to the one that's declared later
+    OutputType = maybe_output_type  # pylint: disable=invalid-name
     # The type that goes in the response is a list of the yielded type
     if get_origin(OutputType) is Iterator:
         # Annotated allows us to attach Field annotations to the list, which we use to mark that this is an iterator
@@ -462,7 +471,10 @@ class TrainingInput(BaseModel):
     train = get_train(predictor)
     signature = inspect.signature(train)
 
-    input_types = get_type_hints(train)
+    try:
+        input_types = get_type_hints(train)
+    except TypeError:
+        input_types = {}
     if "return" in input_types:
         del input_types["return"]
 
@@ -483,8 +495,8 @@ def get_training_output_type(predictor: BasePredictor) -> Type[BaseModel]:
 
     train = get_train(predictor)
 
-    input_types = get_type_hints(train)
-    if "return" not in input_types:
+    TrainingOutputType = get_return_annotation(train)  # pylint: disable=invalid-name
+    if not TrainingOutputType:
         raise TypeError(
             """You must set an output type. If your model can return multiple output types, you can explicitly set `Any` as the output type.
 
@@ -500,8 +512,6 @@ def train(
 """
         )
 
-    TrainingOutputType = input_types.pop("return")  # pylint: disable=invalid-name
-
     name = (
         TrainingOutputType.__name__ if hasattr(TrainingOutputType, "__name__") else ""
     )

test-integration/test_integration/fixtures/partial-predict-project/predict.py

@@ -1,5 +1,6 @@
-from typing import Callable
 import functools
+import inspect
+from typing import Any, Callable
 
 from cog import BasePredictor, Input
 
@@ -10,7 +11,7 @@ def general(
     ) -> int:
         return 1
 
-    def _remove(f: Callable, defaults: dict[str, Any]) -> Callable:
+    def _remove(f: Callable, defaults: "dict[str, Any]") -> Callable:
         # pylint: disable=no-self-argument
         def wrapper(self, *args, **kwargs):
             kwargs.update(defaults)
@@ -31,7 +32,7 @@ def wrapper(self, *args, **kwargs):
     predict = _remove(general, {"system_prompt": ""})
 
 
-def _train(self, prompt: str = Input(description="hi"), system_prompt: str = None):
+def _train(prompt: str = Input(description="hi"), system_prompt: str = None) -> int:
     return 1