replacement_utils

This module provides helper functions for parsing, sorting, and analyzing layer replacement configurations used in the replacement library for model compression.

Functions

`parse_layer_replacement`
`extract_block_configs_and_locations`
`weights_path_to_checkpoint_dir`
`replacement_is_teacher`
`is_replacement_identical_to_teacher`
`split_replacements_to_teacher_and_student`

extract_block_configs_and_locations(layer_replacements)

Parameters:: layer_replacements (list[dict])
Return type:: tuple[list[BlockConfig], list[tuple[dict, int]]]

is_replacement_identical_to_teacher(layer_replacement, teacher_model_config)

Parameters:

layer_replacement (dict)
teacher_model_config (PreTrainedConfig)

Return type:

bool

parse_layer_replacement(layer_replacement)

Parameters:: layer_replacement (dict | str)
Return type:: dict

replacement_is_teacher(layer_replacement, teacher_model_config, teacher_checkpoint_dir)

Parameters:

layer_replacement (dict)
teacher_model_config (PreTrainedConfig)
teacher_checkpoint_dir (Path)

Return type:

bool

split_replacements_to_teacher_and_student(replacements, teacher_model_config, teacher_checkpoint_dir)

Parameters:

replacements (list[dict])
teacher_model_config (PreTrainedConfig)
teacher_checkpoint_dir (Path)

Return type:

tuple[list[dict], list[dict]]

weights_path_to_checkpoint_dir(weights_path)

Parameters:: weights_path (Path)
Return type:: Path