| """Reject stale feature/Markov state after any adapted dependency changes. |
| |
| This is an evaluation guard, checked outside graph capture/replay. Adapters are |
| immutable for a worker's lifetime; a changed dependency requires a fresh worker. |
| """ |
| from __future__ import annotations |
|
|
| from collections.abc import Callable, Mapping |
|
|
| from torch import Tensor |
|
|
|
|
| class DependencyGuard: |
| def __init__(self, dependencies: Mapping[str, Tensor], invalidate: Callable[[], None], |
| current: Callable[[], Mapping[str, Tensor]] | None = None): |
| if not dependencies: |
| raise ValueError("cache guard requires declared adapted dependencies") |
| self.dependencies = dict(dependencies) |
| self.current = current |
| self.versions = {name: (t.data_ptr(), t._version) for name, t in dependencies.items()} |
| self.invalidate = invalidate |
| self.invalidated = False |
|
|
| def check(self) -> None: |
| actual = self.dependencies if self.current is None else self.current() |
| if self.invalidated or actual.keys() != self.versions.keys() or any((t.data_ptr(), t._version) != self.versions[name] |
| for name, t in actual.items()): |
| if not self.invalidated: |
| self.invalidate() |
| self.invalidated = True |
| raise RuntimeError("adapted dependency changed: Markov/feature caches and graphs invalidated; recreate worker") |
|
|
