Skip to navigation
ReferenceFull Library ReferenceNemo AutomodelNemo AutomodelComponentsModelsD1nemo_automodel.components.models.d1.config

nemo_automodel.components.models.d1.config

View as Markdown

Serializable configuration for the released D1 Omni composition.

Module Contents

Classes

NameDescription
D1OmniConfigConfigure the Omni encoder, marker head, and perception modules.

API

class nemo_automodel.components.models.d1.config.D1OmniConfig(
    text_config: dict[str, typing.Any] | transformers.Lfm2Config | None = None,
    vision_config: dict[str, typing.Any] | None = None,
    audio_config: dict[str, typing.Any] | None = None,
    projector_hidden_size: int = 2048,
    head_layers: int = 2,
    max_length: int = 16384,
    image_text_length: int = 896,
    audio_text_length: int = 15360,
    temperatures: dict[str, float] | None = None,
    tie_word_embeddings: bool = False,
    kwargs: typing.Any = {}
)

Bases: PretrainedConfig

Configure the Omni encoder, marker head, and perception modules.

audio_config
= dict(audio_config or {})
model_type
= 'd1_omni'
sub_configs
= {'text_config': Lfm2Config}
temperatures
= dict(temperatures or {})
text_config
vision_config
= dict(vision_config or {})
nemo_automodel.components.models.d1.config.D1OmniConfig.to_dict() -> dict[str, typing.Any]

Serialize HF subconfig fields and the release’s legacy rotary key.

Returns: dict[str, Any]

Configuration mapping reloadable by native and published Omni classes.