cli.merge_lora

cli.merge_lora

CLI to merge a trained LoRA into a base model.

Functions

Name	Description
do_cli	Parses axolotl config, CLI args, and calls do_merge_lora. Note that various
do_merge_lora	Merges LoRA adapters with base model using either memory-efficient or legacy approach.

do_cli

cli.merge_lora.do_cli(
    config=Path('examples/'),
    dequant=False,
    override_quantizer=False,
    **kwargs,
)

Parses axolotl config, CLI args, and calls do_merge_lora. Note that various config values will be overwritten to allow the LoRA merge logic to work as expected (load_in_8bit=False, load_in4bit=False, flash_attention=False, etc.).

Parameters

Name	Type	Description	Default
config	Union[Path, str]	Path to axolotl config YAML file.	Path('examples/')
dequant	bool	If set (--dequant), dequantize a quantized base to bf16 in the merged checkpoint. Default keeps the base’s quantized dtypes (fp8/nvfp4) intact.	False
override_quantizer	bool	If set (--override-quantizer), merge a merge-aware adapter despite a quantizer-identity mismatch (e.g. different torchao version than the one trained with) instead of hard-erroring.	False
kwargs		Additional keyword arguments to override config file values.	{}

Raises

Name	Type	Description
	ValueError	If target directory for LoRA merged model does not exist.

do_merge_lora

cli.merge_lora.do_merge_lora(cfg)

Merges LoRA adapters with base model using either memory-efficient or legacy approach.

Parameters

Name	Type	Description	Default
cfg	DictDefault	Dictionary mapping axolotl config keys to values.	required