MMaDA
/
venv
/lib
/python3.11
/site-packages
/lightning
/fabric
/plugins
/collectives
/collective.py
from abc import ABC, abstractmethod | |
from typing import Any, Optional | |
from torch import Tensor | |
from typing_extensions import Self | |
from lightning.fabric.utilities.types import CollectibleGroup | |
class Collective(ABC): | |
"""Interface for collective operations. | |
Supports communications between multiple processes and multiple nodes. A collective owns a group. | |
.. warning:: This is an :ref:`experimental <versioning:Experimental API>` feature which is still in development. | |
""" | |
def __init__(self) -> None: | |
self._group: Optional[CollectibleGroup] = None | |
def rank(self) -> int: | |
"""Rank.""" | |
def world_size(self) -> int: | |
"""World size.""" | |
def group(self) -> CollectibleGroup: | |
if self._group is None: | |
raise RuntimeError( | |
f"`{type(self).__name__}` does not own a group. HINT: try `collective.create_group().group`" | |
) | |
return self._group | |
def broadcast(self, tensor: Tensor, src: int) -> Tensor: ... | |
def all_reduce(self, tensor: Tensor, op: str) -> Tensor: ... | |
def reduce(self, tensor: Tensor, dst: int, op: str) -> Tensor: ... | |
def all_gather(self, tensor_list: list[Tensor], tensor: Tensor) -> list[Tensor]: ... | |
def gather(self, tensor: Tensor, gather_list: list[Tensor], dst: int = 0) -> list[Tensor]: ... | |
def scatter(self, tensor: Tensor, scatter_list: list[Tensor], src: int = 0) -> Tensor: ... | |
def reduce_scatter(self, output: Tensor, input_list: list[Tensor], op: str) -> Tensor: ... | |
def all_to_all(self, output_tensor_list: list[Tensor], input_tensor_list: list[Tensor]) -> list[Tensor]: ... | |
def send(self, tensor: Tensor, dst: int, tag: int = 0) -> None: ... | |
def recv(self, tensor: Tensor, src: Optional[int] = None, tag: int = 0) -> Tensor: ... | |
def barrier(self, device_ids: Optional[list[int]] = None) -> None: ... | |
def is_available(cls) -> bool: ... | |
def is_initialized(cls) -> bool: ... | |
def init_group(cls, **kwargs: Any) -> None: ... | |
def new_group(cls, **kwargs: Any) -> CollectibleGroup: ... | |
def destroy_group(cls, group: CollectibleGroup) -> None: ... | |
def _convert_to_native_op(cls, op: str) -> Any: ... | |
def setup(self, **kwargs: Any) -> Self: | |
if not self.is_initialized(): | |
self.init_group(**kwargs) | |
return self | |
def create_group(self, **kwargs: Any) -> Self: | |
"""Create a group. | |
This assumes that :meth:`~lightning.fabric.plugins.collectives.Collective.init_group` has been | |
called already by the user. | |
""" | |
if self._group is not None: | |
raise RuntimeError(f"`{type(self).__name__}` already owns a group.") | |
self._group = self.new_group(**kwargs) | |
return self | |
def teardown(self) -> Self: | |
if self._group is None: | |
raise RuntimeError(f"`{type(self).__name__}` does not own a group to destroy.") | |
self.destroy_group(self._group) | |
self._group = None | |
return self | |