nvalchemi.training.peft.lora_hook.LoRAHook#

class nvalchemi.training.peft.lora_hook.LoRAHook(*, lora_config, register_parameters=True)[source]#

Apply LoRA adapters to models.

This hook is automatically prepended when FineTuningStrategy receives a LoRAConfig.

Parameters:
  • lora_config (LoRAConfig) – LoRA PEFT configuration describing adapter targets, rank, scaling, dropout, MLP wrapping, and custom wrapper registrations.

  • register_parameters (bool, optional) – If True, register LoRA adapter parameters as trainable and managed on the workflow. Set to False for standalone model loading where no training workflow registry exists. Defaults to True.

frequency#

Required by the hook protocol; always 1.

Type:

int

stage#

This hook does not run at training stages.

Type:

None

model_config = {'arbitrary_types_allowed': True}#

Configuration for the model, should be a dictionary conforming to [ConfigDict][pydantic.config.ConfigDict].

on_register(workflow)[source]#

Register wrappers, inject adapters, and register adapter parameters.

The public configuration is validated before creating this hook. During registration, model-prefixed targets are converted to model-local targets for each model to be compatible with the PhysicsNeMo apply_lora function. For example, "student.model.projection" becomes the PhysicsNeMo target_modules entry "model.projection" for the "student" model. Adapters are then injected, and their trainable and managed parameter names are registered on the workflow.

Parameters:

workflow (Any)

Return type:

None