Multiple dispatch for modern Python.
Dispatch on what actually matters. Stay explicit under ambiguity. Keep methods feeling like methods.

Fast exact-type hot paths, callable guards, keyword-aware calls, and class fallback that still behaves like Python.

Most Python dispatch tools stop at "pick an implementation based on the first argument".

That is useful, but a lot of real systems need more:

• a compiler that dispatches on (node, backend)
• a renderer that dispatches on (value, target_format)
• a pricing engine that dispatches on (instrument, market)
• a rules engine that dispatches on type first, then value constraints
• a class hierarchy where subclasses specialize a few cases without breaking normal fallback

That is where multimethods lives.

It gives you:

• multiple dispatch across more than one argument
• explicit ambiguity errors instead of silent guesswork
• callable guards instead of string eval
• keyword-call support through canonical signature binding
• @staticmethod and @classmethod support
• ordinary MRO fallback when a subclass only overloads part of the surface
• a fast exact-type cache for repeated hot-path calls

python -m pip install multimethods

Requires Python 3.12+.

This is the shape of the library:

from multimethods import multimethod


@multimethod
def render(value: object, target: object) -> str:
    return "fallback"


@render.register
def _(value: int, target: str) -> str:
    return f"int:{value}:{target}"


@render.register
def _(value: str, target: str) -> str:
    return f"str:{value}:{target}"


assert render(3, "cli") == "int:3:cli"
assert render("hello", "cli") == "str:hello:cli"
assert render(3.14, "cli") == "fallback"

The mental model is simple:

1. write a normal function
2. register overloads
3. let runtime types decide which one wins
4. get an explicit error if the winner is genuinely unclear

A lot of dispatch problems are naturally two-dimensional.

from dataclasses import dataclass

from multimethods import multimethod


@dataclass
class JSON:
    pass


@dataclass
class CSV:
    pass


@dataclass
class User:
    name: str
    email: str


@dataclass
class Invoice:
    total: int
    currency: str


@multimethod
def dump(value: User, target: JSON) -> str:
    return f'{{"name": "{value.name}", "email": "{value.email}"}}'


@dump.register
def _(value: User, target: CSV) -> str:
    return f"{value.name},{value.email}"


@dump.register
def _(value: Invoice, target: JSON) -> str:
    return f'{{"total": {value.total}, "currency": "{value.currency}"}}'

That reads like the problem statement. No manual matrix dispatch table, no nested if isinstance(...) ladders.

Sometimes "type" is not enough. You want int, but only in a specific value range.

from multimethods import multimethod


@multimethod(int, guard=lambda status: status == 429)
def retry_policy(status: int) -> str:
    return "retry-with-backoff"


@retry_policy.register(int, guard=lambda status: 500 <= status < 600)
def _(status: int) -> str:
    return "retry"


@retry_policy.register(int)
def _(status: int) -> str:
    return "do-not-retry"


assert retry_policy(429) == "retry-with-backoff"
assert retry_policy(503) == "retry"
assert retry_policy(404) == "do-not-retry"

This is value-aware dispatch without resorting to string expressions or a second ad hoc rule system.

Sometimes only the first one or two arguments should participate in dispatch, while the rest are normal parameters.

from multimethods import multimethod


@multimethod(int)
def parse(value, *, base=10):
    return value


@parse.register(str)
def _(value, *, base=10):
    return int(value, base=base)


assert parse(12) == 12
assert parse("1111", base=2) == 15

That gives you a clean way to say: "dispatch on the input shape, but keep configuration keyword-only".

Subclass-specific overloads should not destroy ordinary class fallback.

from multimethods import multimethod


class BaseFormatter:
    def format(self, value):
        return str(value)


class ReportFormatter(BaseFormatter):
    @multimethod
    def format(self, value: int):
        return f"{value:,}"


formatter = ReportFormatter()

assert formatter.format(1250000) == "1,250,000"
assert formatter.format("raw") == "raw"

When no local overload matches, multimethods walks the class MRO and calls the next attribute with the same name. That makes partial specialization practical.

For parameter-local rules, where(...) composes cleanly with Annotated.

from typing import Annotated

from multimethods import multimethod, where


@multimethod
def bucket(x: Annotated[int, where(lambda x: x >= 0)]) -> str:
    return "non-negative"


@bucket.register
def _(x: int) -> str:
    return "negative"


assert bucket(10) == "non-negative"
assert bucket(-3) == "negative"

Many dispatch systems quietly choose one candidate when two overloads are both plausible. That makes debugging miserable.

multimethods does not do that.

from multimethods import AmbiguousDispatchError, multimethod


@multimethod
def collide(x: int, y: object):
    return "left"


@collide.register
def _(x: object, y: int):
    return "right"


try:
    collide(1, 1)
except AmbiguousDispatchError:
    print("good: the dispatcher refused to guess")

If you really want one side to win, make that decision explicit with priority=:

@multimethod(priority=10)
def choose(x: int, y: object):
    return "left"


@choose.register(priority=1)
def _(x: object, y: int):
    return "right"

Dispatch works in this order:

1. Find overloads whose type constraints match the runtime dispatch arguments.
2. Remove strictly less specific candidates.
3. Evaluate parameter guards and callable guards.
4. Prefer guarded candidates over unguarded candidates with identical type constraints.
5. Break remaining ties with priority=.
6. Raise AmbiguousDispatchError if multiple winners remain.
7. If this is a bound method call and no local overload matches, walk the class MRO and call the next attribute with the same name.

Keyword-only parameters and variadic parameters are allowed, but they never participate in dispatch.

• plain classes
• abstract base classes
• typing.Any
• unions of supported classes, including int | str
• typing.Annotated[T, where(...)]
• annotation-based registration
• explicit decorator registration like @multimethod(int, str)

multimethods is built for code that wants richer semantics than a minimal dispatch helper, while still caring about performance.

The implementation currently optimizes:

• repeated exact-type calls
• pure type-only overload sets
• positional hot paths that can bypass inspect.Signature.bind

On the included local benchmark, warmed exact-type calls are currently faster than the older multimethod / multidispatch implementations, roughly in the same band as plum, and still behind multipledispatch and especially ovld on raw microbenchmarks.

That is the honest position today:

• strong semantics
• practical ergonomics
• competitive hot-path speed
• still room to push harder on generated fast paths

You can run the benchmark yourself:

python benchmarks/compare_dispatch.py

multimethods is a good fit when your codebase has:

• AST or IR transforms
• serializers and renderers with multiple target formats
• geometry, simulation, or collision-style double dispatch
• pricing or analytics logic driven by pairs of domain objects
• plugin systems where both the payload and backend matter
• business rules that are cleaner as overloads than as branching trees

If a plain if statement is enough, use a plain if statement.

If first-argument dispatch is enough, functools.singledispatch is still excellent.

This library is for the cases after that.

Python has several dispatch options. multimethods targets the gap after functools.singledispatch: multi-argument problems that still need predictable semantics and ordinary method behavior.

When singledispatch is enough: one dispatch axis, stdlib-only, excellent ergonomics. Stay there.

When multipledispatch is tempting: broad adoption and raw speed, but ambiguity handling and OOP fallback are weaker for production rule systems.

When plum is tempting: richest typing integration (parametric types, beartype-powered hints). Choose it when you want that breadth and accept the dependency.

When ovld is tempting: fastest microbenchmarks and a pattern-matching style API. Choose it when codegen speed matters more than Julia-like ambiguity rules.

When multimethods fits: you want zero-dependency multiple dispatch with explicit ambiguity errors, callable guards, and subclass MRO fallback without adopting a larger typing framework.

Every MultiMethod exposes three helpers for debugging, tooling, and extension.

Resolve the winning overload without calling it. Returns the underlying function.

from multimethods import multimethod


@multimethod
def convert(value: object) -> str:
    return str(value)


@convert.register
def _(value: int) -> str:
    return f"int:{value}"


fn = convert.dispatch(3)
assert fn is convert.registry[1].function
assert fn(3) == "int:3"
assert convert(3) == "int:3"

Read-only view of registered overloads after pending forward references resolve.

assert len(convert.registry) == 2
assert all(hasattr(entry, "constraints") for entry in convert.registry)

Clone the dispatcher and its overload table. Caches start empty on the clone.

clone = convert.copy()
assert clone is not convert
assert len(clone.registry) == len(convert.registry)
assert clone(3) == convert(3)

Runtime dispatch is dynamic; static type checkers need explicit stubs. Pair @multimethod overloads with typing.overload declarations that share the same name and parameter shapes.

from typing import overload

from multimethods import multimethod


@overload
def stringify(value: int) -> str: ...


@overload
def stringify(value: str) -> str: ...


@multimethod
def stringify(value: object) -> str:
    return str(value)


@stringify.register
def _(value: int) -> str:
    return f"int:{value}"


@stringify.register
def _(value: str) -> str:
    return f"str:{value}"

mypy and pyright use the @overload stubs for call-site checking; Python uses the @multimethod registrations at runtime. Keep stub signatures aligned with the dispatch parameters of each registered overload.

Unsupported runtime annotations such as Literal[...], list[int], and Protocol are rejected at registration time. Use plain classes plus guard= or Annotated[..., where(...)] for value refinement instead.

python -m pip install -e .[dev]
python -m ruff check .
python -m pytest -q
python -m build
python benchmarks/compare_dispatch.py