nemo_automodel.shared.peft_config | NVIDIA NeMo AutoModel

PEFT configuration without adapter or distributed runtime dependencies.

Module Contents

Classes

Name	Description
`PeftConfig`	Declarative adapter settings shared by execution and checkpoint boundaries.

API

class nemo_automodel.shared.peft_config.PeftConfig(
    target_modules: list[str] | str | None = list(),
    exclude_modules: list[str] | str | None = list(),
    match_all_linear: bool = False,
    dim: int = 8,
    alpha: int = 32,
    use_dora: bool = False,
    dropout: float = 0.0,
    dropout_position: typing.Literal['pre', 'post'] = 'post',
    lora_A_init: str = 'xavier',
    lora_dtype: torch.dtype | str | None = None,
    use_memory_efficient_lora: bool = True,
    use_triton: bool = False,
    moe_rank_scaling: bool = False,
    expert_weight_format: typing.Literal['unquantized', 'mxfp4'] = 'unquantized'
)

Dataclass

Declarative adapter settings shared by execution and checkpoint boundaries.

alpha

int = 32

dim

int = 8

dropout

float = 0.0

dropout_position

Literal['pre', 'post'] = 'post'

exclude_modules

list[str] | str | None = field(default_factory=list)

expert_weight_format

Literal['unquantized', 'mxfp4'] = 'unquantized'

lora_A_init

str = 'xavier'

lora_dtype

dtype | str | None = None

match_all_linear

bool = False

moe_rank_scaling

bool = False

target_modules

list[str] | str | None = field(default_factory=list)

use_dora

bool = False

use_memory_efficient_lora

bool = True

use_triton

bool = False