nvalchemi.training.hooks.ModulePatchHook

pydantic model nvalchemi.training.hooks.ModulePatchHook

Patch model submodules at registration time.

Patches run when the hook is registered on a workflow. Each target path must include the model key followed by an existing parent path and a final child attribute, for example "main.model.projection". The parent module must exist; the final child is added when missing or replaced when present. Shape compatibility is intentionally user-owned and is validated naturally by the model’s forward pass or downstream checkpoint loading.

Warns:

UserWarning – If the same direct module instance is assigned to multiple targets.

Examples

>>> import torch
>>> from nvalchemi.training.hooks import ModulePatchHook
>>> hook = ModulePatchHook(
...     patches={"main.model.projection": torch.nn.Linear(8, 1)}
... )
>>> hook.frequency
1

field patches: dict[str, BaseSpec | Module] [Optional]

Ordered mapping of target paths to replacement modules or specs that build modules.