library#

Replacement library for virtual candidates backed by one sorted teacher checkpoint.

Classes

ReplacementLibrary

class ReplacementLibrary#

Bases: object

__init__(replacement_library_path, descriptor, model_config_overrides=None, sorted_teacher_dir=None)#
Parameters:

  • replacement_library_path (str | Path)

  • model_config_overrides (dict | None)

  • sorted_teacher_dir (str | Path | None)

create_model_config(layer_replacements)#
Parameters:

layer_replacements (list[dict])

get_arbitrary_checkpoint_dir()#
Return type:

Path

load_model(layer_replacements)#

Build a variant model by slicing/merging from the sorted teacher.

Parameters:

layer_replacements (list[dict])

materialize_checkpoint(layer_replacements, output_dir, *, model_config=None, overwrite=False, solution_identity=None)#

Stream a realized child checkpoint without constructing a dense model.

Parameters:

  • layer_replacements (list[dict])

  • output_dir (str | Path)

  • model_config (PreTrainedConfig | None)

  • overwrite (bool)

  • solution_identity (str | None)

Return type:

Path

property model_config: PreTrainedConfig#