Fine-tuning API

Registration-time helpers for adapting pretrained models before optimizer construction.

See also

• User guide: Fine-Tuning Pretrained Models

• Training strategy API: Training strategy API

• Training update hooks: Training update hooks

Strategy

FineTuningStrategy	Training strategy for patching modules and selecting trainable parameters.
FineTuningStrategy.from_pretrained_checkpoint	Start a new fine-tuning run from checkpointed model weights.
FineTuningStrategy.load_checkpoint	Load a restartable strategy checkpoint.

Use FineTuningStrategy.load_checkpoint(...) to resume an interrupted run with saved optimizer state, scheduler state, counters, and serialized fine-tuning configuration. Use FineTuningStrategy.from_pretrained_checkpoint(...) to start a new fine-tuning run whose model weights are initialized from an existing checkpoint; optimizer state, hooks, and counters do not carry over. See Fine-Tuning Pretrained Models for patterns and examples.

Hooks

Registration-time hooks that adapt the model tree and optimizer parameter set before training starts. They do not own backward() or optimizer-step behavior; use Training update hooks for batch-update policies.

ModulePatchHook	Patch model submodules at registration time.
TrainableParameterHook	Select which parameters stay trainable during fine-tuning.

ModulePatchHook

Field	Type	Description
patches	dict[str, nvalchemi.training._spec.BaseSpec | torch.nn.modules.module.Module]	Ordered mapping of target paths to replacement modules or specs that build modules.

TrainableParameterHook

Field	Type	Description
freeze_patterns	tuple[str, ...]	Glob patterns for parameters to freeze. Overridden by trainable_patterns — a parameter matching both stays trainable.
trainable_patterns	tuple[str, ...]	Glob patterns for parameters to keep trainable. On their own they define the full trainable set (everything else is frozen).
freeze_mode	typing.Literal['requires_grad', 'optimizer_only']	Whether excluded parameters are temporarily frozen via requires_grad=False or only excluded from optimizer construction.