Files
pytorch/torch/distributed/_composable_state.py
PyTorch MergeBot d2494cbb2b Revert "[distributed] Replace assert statements with AssertionError exceptions (#165216)"
This reverts commit 74db92b21868b7e9e77cc966e5d57a8246723cbd.

Reverted https://github.com/pytorch/pytorch/pull/165216 on behalf of https://github.com/clee2000 due to I think this broke distributed/test_pg_wrapper.py::ProcessGroupNCCLWrapperTest::test_debug_level_detail_no_gloo [GH job link](https://github.com/pytorch/pytorch/actions/runs/18492765290/job/52693842750) [HUD commit link](74db92b218), note to self: bad TD ([comment](https://github.com/pytorch/pytorch/pull/165216#issuecomment-3402838765))
2025-10-14 17:05:16 +00:00

46 lines
1.4 KiB
Python

import weakref
from typing import cast, Optional
import torch.nn as nn
class _State:
pass
_module_state_mapping: weakref.WeakKeyDictionary[
nn.Module, weakref.ReferenceType[_State]
] = weakref.WeakKeyDictionary()
def _insert_module_state(module: nn.Module, state: _State) -> None:
global _module_state_mapping
assert module not in _module_state_mapping, f"Inserting {module} more than once."
_module_state_mapping[module] = weakref.ref(state)
def _get_module_state(module: nn.Module) -> Optional[_State]:
"""
Return the ``_State`` in ``model``.
Given a ``module``, this API finds out if the module is also a ``_State``
instance or if the module is managed by a composable API. If the module
is also a ``_State``, ``module`` will be casted to ``_State` and returned.
If it is managed by a composable API, the corresponding ``_State`` will
be returned.
"""
global _module_state_mapping
if isinstance(module, _State):
# pyrefly: ignore # redundant-cast
return cast(_State, module)
else:
# https://github.com/pytorch/pytorch/issues/107054
if module in _module_state_mapping:
state_ref = _module_state_mapping[module]
state = state_ref()
if state is None:
raise AssertionError("State has already been garbage collected")
return state
else:
return None