child_init
Core logic for creating pruned child model state dicts from parent models. Used by init_child_from_parent.
Functions
- create_child_state_dict(pruning_mixin, descriptor, original_state_dict, new_state_dict, original_config, new_config, gqa_init_mode, ignore_fn=<function <lambda>>, mlp_init_mode=MlpInitMode.CopyAsIs, mlp_init_config=None, owned_block_indexes=None, linear_init_mode=LinearInitMode.Random, hidden_size_init_mode=HiddenSizeInitMode.CopyAsIs, channel_importance_path=None, max_layer_workers=None)
- Parameters:
original_state_dict (dict)
new_state_dict (dict)
original_config (PreTrainedConfig)
new_config (PreTrainedConfig)
gqa_init_mode (GQAInitMode)
ignore_fn (Callable[[str], bool])
mlp_init_mode (MlpInitMode)
mlp_init_config (dict[str, Any] | None)
owned_block_indexes (set[int] | None)
linear_init_mode (LinearInitMode)
hidden_size_init_mode (HiddenSizeInitMode)
channel_importance_path (str | None)
max_layer_workers (int | None)
- update_model_config(model_config, model_config_overrides=None)
- Parameters:
model_config (PreTrainedConfig)
model_config_overrides (None | list[dict[str, Any]] | str | dict | Path)
- Return type:
PreTrainedConfig