Skip to content

Latest commit

 

History

History
1437 lines (1005 loc) · 44.9 KB

generics.rst

File metadata and controls

1437 lines (1005 loc) · 44.9 KB
Raw

Generics

This section explains how you can define your own generic classes that take one or more type arguments, similar to built-in types such as list[T]. User-defined generics are a moderately advanced feature and you can get far without ever using them -- feel free to skip this section and come back later.

Defining generic classes

The built-in collection classes are generic classes. Generic types accept one or more type arguments within [...], which can be arbitrary types. For example, the type dict[int, str] has the type arguments int and str, and list[int] has the type argument int.

Programs can also define new generic classes. Here is a very simple generic class that represents a stack (using the syntax introduced in Python 3.12):

class Stack[T]:
    def __init__(self) -> None:
        # Create an empty list with items of type T
        self.items: list[T] = []

    def push(self, item: T) -> None:
        self.items.append(item)

    def pop(self) -> T:
        return self.items.pop()

    def empty(self) -> bool:
        return not self.items

There are two syntax variants for defining generic classes in Python. Python 3.12 introduced a new dedicated syntax for defining generic classes (and also functions and type aliases, which we will discuss later). The above example used the new syntax. Most examples are given using both the new and the old (or legacy) syntax variants. Unless mentioned otherwise, they work the same -- but the new syntax is more readable and more convenient.

Here is the same example using the old syntax (required for Python 3.11 and earlier, but also supported on newer Python versions):

from typing import TypeVar, Generic

T = TypeVar('T')  # Define type variable "T"

class Stack(Generic[T]):
    def __init__(self) -> None:
        # Create an empty list with items of type T
        self.items: list[T] = []

    def push(self, item: T) -> None:
        self.items.append(item)

    def pop(self) -> T:
        return self.items.pop()

    def empty(self) -> bool:
        return not self.items

Note

There are currently no plans to deprecate the legacy syntax. You can freely mix code using the new and old syntax variants, even within a single file (but not within a single class).

The Stack class can be used to represent a stack of any type: Stack[int], Stack[tuple[int, str]], etc. You can think of Stack[int] as referring to the definition of Stack above, but with all instances of T replaced with int.

Using Stack is similar to built-in container types:

# Construct an empty Stack[int] instance
stack = Stack[int]()
stack.push(2)
stack.pop()

# error: Argument 1 to "push" of "Stack" has incompatible type "str"; expected "int"
stack.push('x')

stack2: Stack[str] = Stack()
stack2.append('x')

Construction of instances of generic types is type checked (Python 3.12 syntax):

class Box[T]:
    def __init__(self, content: T) -> None:
        self.content = content

Box(1)       # OK, inferred type is Box[int]
Box[int](1)  # Also OK

# error: Argument 1 to "Box" has incompatible type "str"; expected "int"
Box[int]('some string')

Here is the definition of Box using the legacy syntax (Python 3.11 and earlier):

from typing import TypeVar, Generic

T = TypeVar('T')

class Box(Generic[T]):
    def __init__(self, content: T) -> None:
        self.content = content

Note

Before moving on, let's clarify some terminology. The name T in class Stack[T] or class Stack(Generic[T]) declares a type parameter T (of class Stack). T is also called a type variable, especially in a type annotation, such as in the signature of push above. When the type Stack[...] is used in a type annotation, the type within square brackets is called a type argument. This is similar to the distinction between function parameters and arguments.

Defining subclasses of generic classes

User-defined generic classes and generic classes defined in :py:mod:`typing` can be used as a base class for another class (generic or non-generic). For example (Python 3.12 syntax):

from typing import Mapping, Iterator

# This is a generic subclass of Mapping
class MyMap[KT, VT](Mapping[KT, VT]):
    def __getitem__(self, k: KT) -> VT: ...
    def __iter__(self) -> Iterator[KT]: ...
    def __len__(self) -> int: ...

items: MyMap[str, int]  # OK

# This is a non-generic subclass of dict
class StrDict(dict[str, str]):
    def __str__(self) -> str:
        return f'StrDict({super().__str__()})'

data: StrDict[int, int]  # Error! StrDict is not generic
data2: StrDict  # OK

# This is a user-defined generic class
class Receiver[T]:
    def accept(self, value: T) -> None: ...

# This is a generic subclass of Receiver
class AdvancedReceiver[T](Receiver[T]): ...

Here is the above example using the legacy syntax (Python 3.11 and earlier):

from typing import Generic, TypeVar, Mapping, Iterator

KT = TypeVar('KT')
VT = TypeVar('VT')

# This is a generic subclass of Mapping
class MyMap(Mapping[KT, VT]):
    def __getitem__(self, k: KT) -> VT: ...
    def __iter__(self) -> Iterator[KT]: ...
    def __len__(self) -> int: ...

items: MyMap[str, int]  # OK

# This is a non-generic subclass of dict
class StrDict(dict[str, str]):
    def __str__(self) -> str:
        return f'StrDict({super().__str__()})'

data: StrDict[int, int]  # Error! StrDict is not generic
data2: StrDict  # OK

# This is a user-defined generic class
class Receiver(Generic[T]):
    def accept(self, value: T) -> None: ...

# This is a generic subclass of Receiver
class AdvancedReceiver(Receiver[T]): ...

Note

You have to add an explicit :py:class:`~collections.abc.Mapping` base class if you want mypy to consider a user-defined class as a mapping (and :py:class:`~collections.abc.Sequence` for sequences, etc.). This is because mypy doesn't use structural subtyping for these ABCs, unlike simpler protocols like :py:class:`~collections.abc.Iterable`, which use :ref:`structural subtyping <protocol-types>`.

When using the legacy syntax, :py:class:`Generic <typing.Generic>` can be omitted from bases if there are other base classes that include type variables, such as Mapping[KT, VT] in the above example. If you include Generic[...] in bases, then it should list all type variables present in other bases (or more, if needed). The order of type parameters is defined by the following rules:

  • If Generic[...] is present, then the order of parameters is always determined by their order in Generic[...].
  • If there are no Generic[...] in bases, then all type parameters are collected in the lexicographic order (i.e. by first appearance).

Example:

from typing import Generic, TypeVar, Any

T = TypeVar('T')
S = TypeVar('S')
U = TypeVar('U')

class One(Generic[T]): ...
class Another(Generic[T]): ...

class First(One[T], Another[S]): ...
class Second(One[T], Another[S], Generic[S, U, T]): ...

x: First[int, str]        # Here T is bound to int, S is bound to str
y: Second[int, str, Any]  # Here T is Any, S is int, and U is str

When using the Python 3.12 syntax, all type parameters must always be explicitly defined immediately after the class name within [...], and the Generic[...] base class is never used.

Generic functions

Functions can also be generic, i.e. they can have type parameters (Python 3.12 syntax):

from collections.abc import Sequence

# A generic function!
def first[T](seq: Sequence[T]) -> T:
    return seq[0]

Here is the same example using the legacy syntax (Python 3.11 and earlier):

from typing import TypeVar, Sequence

T = TypeVar('T')

# A generic function!
def first(seq: Sequence[T]) -> T:
    return seq[0]

As with generic classes, the type parameter T can be replaced with any type. That means first can be passed an argument with any sequence type, and the return type is derived from the sequence item type. Example:

reveal_type(first([1, 2, 3]))   # Revealed type is "builtins.int"
reveal_type(first(('a', 'b')))  # Revealed type is "builtins.str"

When using the legacy syntax, a single definition of a type variable (such as T above) can be used in multiple generic functions or classes. In this example we use the same type variable in two generic functions to declarare type parameters:

from typing import TypeVar, Sequence

T = TypeVar('T')      # Define type variable

def first(seq: Sequence[T]) -> T:
    return seq[0]

def last(seq: Sequence[T]) -> T:
    return seq[-1]

Since the Python 3.12 syntax is more concise, it doesn't need (or have) an equivalent way of sharing type parameter definitions.

A variable cannot have a type variable in its type unless the type variable is bound in a containing generic class or function.

When calling a generic function, you can't explicitly pass the values of type parameters as type arguments. The values of type parameters are always inferred by mypy. This is not valid:

first[int]([1, 2])  # Error: can't use [...] with generic function

If you really need this, you can define a generic class with a __call__ method.

Type variables with upper bounds

A type variable can also be restricted to having values that are subtypes of a specific type. This type is called the upper bound of the type variable, and it is specified using T: <bound> when using the Python 3.12 syntax. In the definition of a generic function or a generic class that uses such a type variable T, the type represented by T is assumed to be a subtype of its upper bound, so you can use methods of the upper bound on values of type T (Python 3.12 syntax):

from typing import SupportsAbs

def max_by_abs[T: SupportsAbs[float]](*xs: T) -> T:
    # We can use abs(), because T is a subtype of SupportsAbs[float].
    return max(xs, key=abs)

An upper bound can also be specified with the bound=... keyword argument to :py:class:`~typing.TypeVar`. Here is the example using the legacy syntax (Python 3.11 and earlier):

from typing import TypeVar, SupportsAbs

T = TypeVar('T', bound=SupportsAbs[float])

def max_by_abs(*xs: T) -> T:
    return max(xs, key=abs)

In a call to such a function, the type T must be replaced by a type that is a subtype of its upper bound. Continuing the example above:

max_by_abs(-3.5, 2)   # Okay, has type 'float'
max_by_abs(5+6j, 7)   # Okay, has type 'complex'
max_by_abs('a', 'b')  # Error: 'str' is not a subtype of SupportsAbs[float]

Type parameters of generic classes may also have upper bounds, which restrict the valid values for the type parameter in the same way.

Generic methods and generic self

You can also define generic methods. In particular, the self parameter may also be generic, allowing a method to return the most precise type known at the point of access. In this way, for example, you can type check a chain of setter methods (Python 3.12 syntax):

class Shape:
    def set_scale[T: Shape](self: T, scale: float) -> T:
        self.scale = scale
        return self

class Circle(Shape):
    def set_radius(self, r: float) -> 'Circle':
        self.radius = r
        return self

class Square(Shape):
    def set_width(self, w: float) -> 'Square':
        self.width = w
        return self

circle: Circle = Circle().set_scale(0.5).set_radius(2.7)
square: Square = Square().set_scale(0.5).set_width(3.2)

Without using generic self, the last two lines could not be type checked properly, since the return type of set_scale would be Shape, which doesn't define set_radius or set_width.

When using the legacy syntax, just use a type variable in the method signature that is different from class type parameters (if any are defined). Here is the above example using the legacy syntax (3.11 and earlier):

from typing import TypeVar

T = TypeVar('T', bound='Shape')

class Shape:
    def set_scale(self: T, scale: float) -> T:
        self.scale = scale
        return self

class Circle(Shape):
    def set_radius(self, r: float) -> 'Circle':
        self.radius = r
        return self

class Square(Shape):
    def set_width(self, w: float) -> 'Square':
        self.width = w
        return self

circle: Circle = Circle().set_scale(0.5).set_radius(2.7)
square: Square = Square().set_scale(0.5).set_width(3.2)

Other uses include factory methods, such as copy and deserialization methods. For class methods, you can also define generic cls, using type[T] or :py:class:`Type[T] <typing.Type>` (Python 3.12 syntax):

class Friend:
    other: "Friend | None" = None

    @classmethod
    def make_pair[T: Friend](cls: type[T]) -> tuple[T, T]:
        a, b = cls(), cls()
        a.other = b
        b.other = a
        return a, b

class SuperFriend(Friend):
    pass

a, b = SuperFriend.make_pair()

Here is the same example using the legacy syntax (3.11 and earlier):

from typing import TypeVar

T = TypeVar('T', bound='Friend')

class Friend:
    other: "Friend | None" = None

    @classmethod
    def make_pair(cls: type[T]) -> tuple[T, T]:
        a, b = cls(), cls()
        a.other = b
        b.other = a
        return a, b

class SuperFriend(Friend):
    pass

a, b = SuperFriend.make_pair()

Note that when overriding a method with generic self, you must either return a generic self too, or return an instance of the current class. In the latter case, you must implement this method in all future subclasses.

Note also that mypy cannot always verify that the implementation of a copy or a deserialization method returns the actual type of self. Therefore you may need to silence mypy inside these methods (but not at the call site), possibly by making use of the Any type or a # type: ignore comment.

Mypy lets you use generic self types in certain unsafe ways in order to support common idioms. For example, using a generic self type in an argument type is accepted even though it's unsafe (Python 3.12 syntax):

class Base:
    def compare[T: Base](self: T, other: T) -> bool:
        return False

class Sub(Base):
    def __init__(self, x: int) -> None:
        self.x = x

    # This is unsafe (see below) but allowed because it's
    # a common pattern and rarely causes issues in practice.
    def compare(self, other: 'Sub') -> bool:
        return self.x > other.x

b: Base = Sub(42)
b.compare(Base())  # Runtime error here: 'Base' object has no attribute 'x'

For some advanced uses of self types, see :ref:`additional examples <advanced_self>`.

Automatic self types using typing.Self

Since the patterns described above are quite common, mypy supports a simpler syntax, introduced in PEP 673, to make them easier to use. Instead of introducing a type parameter and using an explicit annotation for self, you can import the special type typing.Self that is automatically transformed into a method-level type parameter with the current class as the upper bound, and you don't need an annotation for self (or cls in class methods). The example from the previous section can be made simpler by using Self:

from typing import Self

class Friend:
    other: Self | None = None

    @classmethod
    def make_pair(cls) -> tuple[Self, Self]:
        a, b = cls(), cls()
        a.other = b
        b.other = a
        return a, b

class SuperFriend(Friend):
    pass

a, b = SuperFriend.make_pair()

This is more compact than using explicit type parameters. Also, you can use Self in attribute annotations in addition to methods.

Note

To use this feature on Python versions earlier than 3.11, you will need to import Self from typing_extensions (version 4.0 or newer).

Variance of generic types

There are three main kinds of generic types with respect to subtype relations between them: invariant, covariant, and contravariant. Assuming that we have a pair of types A and B, and B is a subtype of A, these are defined as follows:

  • A generic class MyCovGen[T] is called covariant in type variable T if MyCovGen[B] is always a subtype of MyCovGen[A].
  • A generic class MyContraGen[T] is called contravariant in type variable T if MyContraGen[A] is always a subtype of MyContraGen[B].
  • A generic class MyInvGen[T] is called invariant in T if neither of the above is true.

Let us illustrate this by few simple examples:

# We'll use these classes in the examples below
class Shape: ...
class Triangle(Shape): ...
class Square(Shape): ...

  • Most immutable container types, such as :py:class:`~collections.abc.Sequence` and :py:class:`~frozenset` are covariant. Union types are also covariant in all union items: Triangle | int is a subtype of Shape | int.

    def count_lines(shapes: Sequence[Shape]) -> int:
    return sum(shape.num_sides for shape in shapes)

triangles: Sequence[Triangle]
count_lines(triangles)  # OK

def foo(triangle: Triangle, num: int) -> None:
    shape_or_number: Union[Shape, int]
    # a Triangle is a Shape, and a Shape is a valid Union[Shape, int]
    shape_or_number = triangle

    Covariance should feel relatively intuitive, but contravariance and invariance can be harder to reason about.

  • :py:class:`~collections.abc.Callable` is an example of type that behaves contravariant in types of arguments. That is, Callable[[Shape], int] is a subtype of Callable[[Triangle], int], despite Shape being a supertype of Triangle. To understand this, consider:

    def cost_of_paint_required(
    triangle: Triangle,
    area_calculator: Callable[[Triangle], float]
) -> float:
    return area_calculator(triangle) * DOLLAR_PER_SQ_FT

# This straightforwardly works
def area_of_triangle(triangle: Triangle) -> float: ...
cost_of_paint_required(triangle, area_of_triangle)  # OK

# But this works as well!
def area_of_any_shape(shape: Shape) -> float: ...
cost_of_paint_required(triangle, area_of_any_shape)  # OK

    cost_of_paint_required needs a callable that can calculate the area of a triangle. If we give it a callable that can calculate the area of an arbitrary shape (not just triangles), everything still works.

  • list is an invariant generic type. Naively, one would think that it is covariant, like :py:class:`~collections.abc.Sequence` above, but consider this code:

    class Circle(Shape):
    # The rotate method is only defined on Circle, not on Shape
    def rotate(self): ...

def add_one(things: list[Shape]) -> None:
    things.append(Shape())

my_circles: list[Circle] = []
add_one(my_circles)     # This may appear safe, but...
my_circles[-1].rotate()  # ...this will fail, since my_circles[0] is now a Shape, not a Circle

    Another example of invariant type is dict. Most mutable containers are invariant.

When using the Python 3.12 syntax for generics, mypy will automatically infer the most flexible variance for each class type variable. Here Box will be inferred as covariant:

class Box[T]:  # this type is implicitly covariant
    def __init__(self, content: T) -> None:
        self._content = content

    def get_content(self) -> T:
        return self._content

def look_into(box: Box[Shape]): ...

my_box = Box(Square())
look_into(my_box)  # OK, but mypy would complain here for an invariant type

Here the underscore prefix for _content is significant. Without an underscore prefix, the class would be invariant, as the attribute would be understood as a public, mutable attribute (a single underscore prefix has no special significance for mypy in most other contexts). By declaring the attribute as Final, the class could still be made covariant:

from typing import Final

class Box[T]:  # this type is implicitly covariant
    def __init__(self, content: T) -> None:
        self.content: Final = content

    def get_content(self) -> T:
        return self.content

When using the legacy syntax, mypy assumes that all user-defined generics are invariant by default. To declare a given generic class as covariant or contravariant, use type variables defined with special keyword arguments covariant or contravariant. For example (Python 3.11 or earlier):

from typing import Generic, TypeVar

T_co = TypeVar('T_co', covariant=True)

class Box(Generic[T_co]):  # this type is declared covariant
    def __init__(self, content: T_co) -> None:
        self._content = content

    def get_content(self) -> T_co:
        return self._content

def look_into(box: Box[Shape]): ...

my_box = Box(Square())
look_into(my_box)  # OK, but mypy would complain here for an invariant type

Type variables with value restriction

By default, a type variable can be replaced with any type -- or any type that is a subtype of the upper bound, which defaults to object. However, sometimes it's useful to have a type variable that can only have some specific types as its value. A typical example is a type variable that can only have values str and bytes. This lets us define a function that can concatenate two strings or bytes objects, but it can't be called with other argument types (Python 3.12 syntax):

def concat[S: (str, bytes)](x: S, y: S) -> S:
    return x + y

concat('a', 'b')    # Okay
concat(b'a', b'b')  # Okay
concat(1, 2)        # Error!

The same thing is also possibly using the legacy syntax (Python 3.11 or earlier):

from typing import TypeVar

AnyStr = TypeVar('AnyStr', str, bytes)

def concat(x: AnyStr, y: AnyStr) -> AnyStr:
    return x + y

No matter which syntax you use, such a type variable is called a type variable with a value restriction. Importantly, this is different from a union type, since combinations of str and bytes are not accepted:

concat('string', b'bytes')   # Error!

In this case, this is exactly what we want, since it's not possible to concatenate a string and a bytes object! If we tried to use a union type, the type checker would complain about this possibility:

def union_concat(x: str | bytes, y: str | bytes) -> str | bytes:
    return x + y  # Error: can't concatenate str and bytes

Another interesting special case is calling concat() with a subtype of str:

class S(str): pass

ss = concat(S('foo'), S('bar'))
reveal_type(ss)  # Revealed type is "builtins.str"

You may expect that the type of ss is S, but the type is actually str: a subtype gets promoted to one of the valid values for the type variable, which in this case is str.

This is thus subtly different from using str | bytes as an upper bound, where the return type would be S (see :ref:`type-variable-upper-bound`). Using a value restriction is correct for concat, since concat actually returns a str instance in the above example:

>>> print(type(ss))
<class 'str'>

You can also use type variables with a restricted set of possible values when defining a generic class. For example, the type :py:class:`Pattern[S] <typing.Pattern>` is used for the return value of :py:func:`re.compile`, where S can be either str or bytes. Regular expressions can be based on a string or a bytes pattern.

A type variable may not have both a value restriction and an upper bound.

Note that you may come across :py:data:`~typing.AnyStr` imported from :py:mod:`typing`. This feature is now deprecated, but it means the same as our definition of AnyStr above.

Declaring decorators

Decorators are typically functions that take a function as an argument and return another function. Describing this behaviour in terms of types can be a little tricky; we'll show how you can use type variables and a special kind of type variable called a parameter specification to do so.

Suppose we have the following decorator, not type annotated yet, that preserves the original function's signature and merely prints the decorated function's name:

def printing_decorator(func):
    def wrapper(*args, **kwds):
        print("Calling", func)
        return func(*args, **kwds)
    return wrapper

We can use it to decorate function add_forty_two:

# A decorated function.
@printing_decorator
def add_forty_two(value: int) -> int:
    return value + 42

a = add_forty_two(3)

Since printing_decorator is not type-annotated, the following won't get type checked:

reveal_type(a)        # Revealed type is "Any"
add_forty_two('foo')  # No type checker error :(

This is a sorry state of affairs! If you run with --strict, mypy will even alert you to this fact: Untyped decorator makes function "add_forty_two" untyped

Note that class decorators are handled differently than function decorators in mypy: decorating a class does not erase its type, even if the decorator has incomplete type annotations.

Here's how one could annotate the decorator (Python 3.12 syntax):

from collections.abc import Callable
from typing import Any, cast

# A decorator that preserves the signature.
def printing_decorator[F: Callable[..., Any]](func: F) -> F:
    def wrapper(*args, **kwds):
        print("Calling", func)
        return func(*args, **kwds)
    return cast(F, wrapper)

@printing_decorator
def add_forty_two(value: int) -> int:
    return value + 42

a = add_forty_two(3)
reveal_type(a)      # Revealed type is "builtins.int"
add_forty_two('x')  # Argument 1 to "add_forty_two" has incompatible type "str"; expected "int"

Here is the example using the legacy syntax (Python 3.11 and earlier):

from collections.abc import Callable
from typing import Any, TypeVar, cast

F = TypeVar('F', bound=Callable[..., Any])

# A decorator that preserves the signature.
def printing_decorator(func: F) -> F:
    def wrapper(*args, **kwds):
        print("Calling", func)
        return func(*args, **kwds)
    return cast(F, wrapper)

@printing_decorator
def add_forty_two(value: int) -> int:
    return value + 42

a = add_forty_two(3)
reveal_type(a)      # Revealed type is "builtins.int"
add_forty_two('x')  # Argument 1 to "add_forty_two" has incompatible type "str"; expected "int"

This still has some shortcomings. First, we need to use the unsafe :py:func:`~typing.cast` to convince mypy that wrapper() has the same signature as func (see :ref:`casts <casts>`).

Second, the wrapper() function is not tightly type checked, although wrapper functions are typically small enough that this is not a big problem. This is also the reason for the :py:func:`~typing.cast` call in the return statement in printing_decorator().

However, we can use a parameter specification, introduced using **P, for a more faithful type annotation (Python 3.12 syntax):

from collections.abc import Callable

def printing_decorator[**P, T](func: Callable[P, T]) -> Callable[P, T]:
    def wrapper(*args: P.args, **kwds: P.kwargs) -> T:
        print("Calling", func)
        return func(*args, **kwds)
    return wrapper

The same is possible using the legacy syntax with :py:class:`~typing.ParamSpec` (Python 3.11 and earlier):

from collections.abc import Callable
from typing import TypeVar
from typing_extensions import ParamSpec

P = ParamSpec('P')
T = TypeVar('T')

def printing_decorator(func: Callable[P, T]) -> Callable[P, T]:
    def wrapper(*args: P.args, **kwds: P.kwargs) -> T:
        print("Calling", func)
        return func(*args, **kwds)
    return wrapper

Parameter specifications also allow you to describe decorators that alter the signature of the input function (Python 3.12 syntax):

from collections.abc import Callable

# We reuse 'P' in the return type, but replace 'T' with 'str'
def stringify[**P, T](func: Callable[P, T]) -> Callable[P, str]:
    def wrapper(*args: P.args, **kwds: P.kwargs) -> str:
        return str(func(*args, **kwds))
    return wrapper

 @stringify
 def add_forty_two(value: int) -> int:
     return value + 42

 a = add_forty_two(3)
 reveal_type(a)      # Revealed type is "builtins.str"
 add_forty_two('x')  # error: Argument 1 to "add_forty_two" has incompatible type "str"; expected "int"

Here is the above example using the legacy syntax (Python 3.11 and earlier):

from collections.abc import Callable
from typing import TypeVar
from typing_extensions import ParamSpec

P = ParamSpec('P')
T = TypeVar('T')

# We reuse 'P' in the return type, but replace 'T' with 'str'
def stringify(func: Callable[P, T]) -> Callable[P, str]:
    def wrapper(*args: P.args, **kwds: P.kwargs) -> str:
        return str(func(*args, **kwds))
    return wrapper

You can also insert an argument in a decorator (Python 3.12 syntax):

from collections.abc import Callable
from typing import Concatenate

def printing_decorator[**P, T](func: Callable[P, T]) -> Callable[Concatenate[str, P], T]:
    def wrapper(msg: str, /, *args: P.args, **kwds: P.kwargs) -> T:
        print("Calling", func, "with", msg)
        return func(*args, **kwds)
    return wrapper

@printing_decorator
def add_forty_two(value: int) -> int:
    return value + 42

a = add_forty_two('three', 3)

Here is the same function using the legacy syntax (Python 3.11 and earlier):

from collections.abc import Callable
from typing import TypeVar
from typing_extensions import Concatenate, ParamSpec

P = ParamSpec('P')
T = TypeVar('T')

def printing_decorator(func: Callable[P, T]) -> Callable[Concatenate[str, P], T]:
    def wrapper(msg: str, /, *args: P.args, **kwds: P.kwargs) -> T:
        print("Calling", func, "with", msg)
        return func(*args, **kwds)
    return wrapper

Decorator factories

Functions that take arguments and return a decorator (also called second-order decorators), are similarly supported via generics (Python 3.12 syntax):

from colletions.abc import Callable
from typing import Any

def route[F: Callable[..., Any]](url: str) -> Callable[[F], F]:
    ...

@route(url='/')
def index(request: Any) -> str:
    return 'Hello world'

Note that mypy infers that F is used to make the Callable return value of route generic, instead of making route itself generic, since F is only used in the return type. Python has no explicit syntax to mark that F is only bound in the return value.

Here is the example using the legacy syntax (Python 3.11 and earlier):

from collections.abc import Callable
from typing import Any, TypeVar

F = TypeVar('F', bound=Callable[..., Any])

def route(url: str) -> Callable[[F], F]:
    ...

@route(url='/')
def index(request: Any) -> str:
    return 'Hello world'

Sometimes the same decorator supports both bare calls and calls with arguments. This can be achieved by combining with :py:func:`@overload <typing.overload>` (Python 3.12 syntax):

from collections.abc import Callable
from typing import Any, overload

# Bare decorator usage
@overload
def atomic[F: Callable[..., Any]](func: F, /) -> F: ...
# Decorator with arguments
@overload
def atomic[F: Callable[..., Any]](*, savepoint: bool = True) -> Callable[[F], F]: ...

# Implementation
def atomic(func: Callable[..., Any] | None = None, /, *, savepoint: bool = True):
    def decorator(func: Callable[..., Any]):
        ...  # Code goes here
    if __func is not None:
        return decorator(__func)
    else:
        return decorator

# Usage
@atomic
def func1() -> None: ...

@atomic(savepoint=False)
def func2() -> None: ...

Here is the decorator from the example using the legacy syntax (Python 3.11 and earlier):

from collections.abc import Callable
from typing import Any, Optional, TypeVar, overload

F = TypeVar('F', bound=Callable[..., Any])

# Bare decorator usage
@overload
def atomic(func: F, /) -> F: ...
# Decorator with arguments
@overload
def atomic(*, savepoint: bool = True) -> Callable[[F], F]: ...

# Implementation
def atomic(func: Optional[Callable[..., Any]] = None, /, *, savepoint: bool = True):
    ...  # Same as above

Generic protocols

Mypy supports generic protocols (see also :ref:`protocol-types`). Several :ref:`predefined protocols <predefined_protocols>` are generic, such as :py:class:`Iterable[T] <collections.abc.Iterable>`, and you can define additional generic protocols. Generic protocols mostly follow the normal rules for generic classes. Example (Python 3.12 syntax):

from typing import Protocol

class Box[T](Protocol):
    content: T

def do_stuff(one: Box[str], other: Box[bytes]) -> None:
    ...

class StringWrapper:
    def __init__(self, content: str) -> None:
        self.content = content

class BytesWrapper:
    def __init__(self, content: bytes) -> None:
        self.content = content

do_stuff(StringWrapper('one'), BytesWrapper(b'other'))  # OK

x: Box[float] = ...
y: Box[int] = ...
x = y  # Error -- Box is invariant

Here is the definition of Box from the above example using the legacy syntax (Python 3.11 and earlier):

from typing import Protocol, TypeVar

T = TypeVar('T')

class Box(Protocol[T]):
    content: T

Note that class ClassName(Protocol[T]) is allowed as a shorthand for class ClassName(Protocol, Generic[T]) when using the legacy syntax, as per :pep:`PEP 544: Generic protocols <544#generic-protocols>`. This form is only valid when using the legacy syntax.

When using the legacy syntax, there is an important difference between generic protocols and ordinary generic classes: mypy checks that the declared variances of generic type variables in a protocol match how they are used in the protocol definition. The protocol in this example is rejected, since the type variable T is used covariantly as a return type, but the type variable is invariant:

from typing import Protocol, TypeVar

T = TypeVar('T')

class ReadOnlyBox(Protocol[T]):  # error: Invariant type variable "T" used in protocol where covariant one is expected
    def content(self) -> T: ...

This example correctly uses a covariant type variable:

from typing import Protocol, TypeVar

T_co = TypeVar('T_co', covariant=True)

class ReadOnlyBox(Protocol[T_co]):  # OK
    def content(self) -> T_co: ...

ax: ReadOnlyBox[float] = ...
ay: ReadOnlyBox[int] = ...
ax = ay  # OK -- ReadOnlyBox is covariant

See :ref:`variance-of-generics` for more about variance.

Generic protocols can also be recursive. Example (Python 3.12 synta):

class Linked[T](Protocol):
    val: T
    def next(self) -> 'Linked[T]': ...

class L:
    val: int
    def next(self) -> 'L': ...

def last(seq: Linked[T]) -> T: ...

result = last(L())
reveal_type(result)  # Revealed type is "builtins.int"

Here is the definition of Linked using the legacy syntax (Python 3.11 and earlier):

from typing import TypeVar

T = TypeVar('T')

class Linked(Protocol[T]):
    val: T
    def next(self) -> 'Linked[T]': ...

Generic type aliases

Type aliases can be generic. In this case they can be used in two ways. First, subscripted aliases are equivalent to original types with substituted type variables. Second, unsubscripted aliases are treated as original types with type parameters replaced with Any.

The type statement introduced in Python 3.12 is used to define generic type aliases (it also supports non-generic type aliases):

from collections.abc import Callable, Iterable

type TInt[S] = tuple[int, S]
type UInt[S] = S | int
type CBack[S] = Callable[..., S]

def response(query: str) -> UInt[str]:  # Same as str | int
    ...
def activate[S](cb: CBack[S]) -> S:        # Same as Callable[..., S]
    ...
table_entry: TInt  # Same as tuple[int, Any]

type Vec[T: (int, float, complex)] = Iterable[tuple[T, T]]

def inproduct[T: (int, float, complex)](v: Vec[T]) -> T:
    return sum(x*y for x, y in v)

def dilate[T: (int, float, complex)](v: Vec[T], scale: T) -> Vec[T]:
    return ((x * scale, y * scale) for x, y in v)

v1: Vec[int] = []      # Same as Iterable[tuple[int, int]]
v2: Vec = []           # Same as Iterable[tuple[Any, Any]]
v3: Vec[int, int] = [] # Error: Invalid alias, too many type arguments!

There is also a legacy syntax that relies on TypeVar. Here the number of type arguments must match the number of free type variables in the generic type alias definition. A type variables is free if it's not a type parameter of a surrounding class or function. Example (following :pep:`PEP 484: Type aliases <484#type-aliases>`, Python 3.11 and earlier):

from typing import TypeVar, Iterable, Union, Callable

S = TypeVar('S')

TInt = tuple[int, S]  # 1 type parameter, since only S is free
UInt = Union[S, int]
CBack = Callable[..., S]

def response(query: str) -> UInt[str]:  # Same as Union[str, int]
    ...
def activate(cb: CBack[S]) -> S:        # Same as Callable[..., S]
    ...
table_entry: TInt  # Same as tuple[int, Any]

T = TypeVar('T', int, float, complex)

Vec = Iterable[tuple[T, T]]

def inproduct(v: Vec[T]) -> T:
    return sum(x*y for x, y in v)

def dilate(v: Vec[T], scale: T) -> Vec[T]:
    return ((x * scale, y * scale) for x, y in v)

v1: Vec[int] = []      # Same as Iterable[tuple[int, int]]
v2: Vec = []           # Same as Iterable[tuple[Any, Any]]
v3: Vec[int, int] = [] # Error: Invalid alias, too many type arguments!

Type aliases can be imported from modules just like other names. An alias can also target another alias, although building complex chains of aliases is not recommended -- this impedes code readability, thus defeating the purpose of using aliases. Example (Python 3.12 syntax):

from example1 import AliasType
from example2 import Vec

# AliasType and Vec are type aliases (Vec as defined above)

def fun() -> AliasType:
    ...

type OIntVec = Vec[int] | None

Type aliases defined using the type statement are not valid as base classes, and they can't be used to construct instances:

from example1 import AliasType
from example2 import Vec

# AliasType and Vec are type aliases (Vec as defined above)

class NewVec[T](Vec[T]):  # Error: not valid as base class
    ...

x = AliasType()  # Error: can't be used to create instances

Here are examples using the legacy syntax (Python 3.11 and earlier):

from typing import TypeVar, Generic, Optional
from example1 import AliasType
from example2 import Vec

# AliasType and Vec are type aliases (Vec as defined above)

def fun() -> AliasType:
    ...

OIntVec = Optional[Vec[int]]

T = TypeVar('T')

# Old-style type aliases can be used as base classes and you can
# construct instances using them

class NewVec(Vec[T]):
    ...

x = AliasType()

for i, j in NewVec[int]():
    ...

Using type variable bounds or value restriction in generic aliases has the same effect as in generic classes and functions.

Differences between the new and old syntax

There are a few notable differences between the new (Python 3.12 and later) and the old syntax for generic classes, functions and type aliases, beyond the obvious syntactic differences:

  • Type variables defined using the old syntax create definitions at runtime in the surrounding namespace, whereas the type variables defined using the new syntax are only defined within the class, function or type variable that uses them.
  • Type variable definitions can be shared when using the old syntax, but the new syntax doesn't support this.
  • When using the new syntax, the variance of class type variables is always inferred.
  • Type aliases defined using the new syntax can contain forward references and recursive references without using string literal escaping. The same is true for the bounds and constraints of type variables.
  • The new syntax lets you define a generic alias where the definition doesn't contain a reference to a type parameter. This is occasionally useful, at least when conditionally defining type aliases.
  • Type aliases defined using the new syntax can't be used as base classes and can't be used to construct instances, unlike aliases defined using the old syntax.

Generic class internals

You may wonder what happens at runtime when you index a generic class. Indexing returns a generic alias to the original class that returns instances of the original class on instantiation (Python 3.12 syntax):

>>> class Stack[T]: ...
>>> Stack
__main__.Stack
>>> Stack[int]
__main__.Stack[int]
>>> instance = Stack[int]()
>>> instance.__class__
__main__.Stack

Here is the same example using the legacy syntax (Python 3.11 and earlier):

>>> from typing import TypeVar, Generic
>>> T = TypeVar('T')
>>> class Stack(Generic[T]): ...
>>> Stack
__main__.Stack
>>> Stack[int]
__main__.Stack[int]
>>> instance = Stack[int]()
>>> instance.__class__
__main__.Stack

Generic aliases can be instantiated or subclassed, similar to real classes, but the above examples illustrate that type variables are erased at runtime. Generic Stack instances are just ordinary Python objects, and they have no extra runtime overhead or magic due to being generic, other than the Generic base class that overloads the indexing operator using __class_getitem__. typing.Generic is included as an implicit base class even when using the new syntax:

>>> class Stack[T]: ...
>>> Stack.mro()
[<class '__main__.Stack'>, <class 'typing.Generic'>, <class 'object'>]

Note that in Python 3.8 and earlier, the built-in types :py:class:`list`, :py:class:`dict` and others do not support indexing. This is why we have the aliases :py:class:`~typing.List`, :py:class:`~typing.Dict` and so on in the :py:mod:`typing` module. Indexing these aliases gives you a generic alias that resembles generic aliases constructed by directly indexing the target class in more recent versions of Python:

>>> # Only relevant for Python 3.8 and below
>>> # If using Python 3.9 or newer, prefer the 'list[int]' syntax
>>> from typing import List
>>> List[int]
typing.List[int]

Note that the generic aliases in typing don't support constructing instances, unlike the corresponding built-in classes:

>>> list[int]()
[]
>>> from typing import List
>>> List[int]()
Traceback (most recent call last):
...
TypeError: Type List cannot be instantiated; use list() instead