nemo_automodel.components.models.muse_spark.configuration_muse_spark

Native runtime configuration for the nested Muse Spark Open Hugging Face release.

Module Contents

Classes

Name	Description
`MuseSparkConfig`	Read/write the nested HF schema while exposing the native flat text contract.

API

class nemo_automodel.components.models.muse_spark.configuration_muse_spark.MuseSparkConfig(
    text_config: dict[str, typing.Any] | None = None,
    vision_config: dict[str, typing.Any] | None = None,
    audio_config: dict[str, typing.Any] | None = None,
    out_hidden_size: int | None = None,
    projector_hidden_size: int = 4096,
    projector_hidden_act: str = 'gelu',
    image_token_id: int = 200092,
    video_token_id: int = 200091,
    audio_token_id: int = 200038,
    kwargs: typing.Any = {}
)

Bases: MuseSpark1xOmniConfig

Read/write the nested HF schema while exposing the native flat text contract.

The native backbone and distributed machinery consume the inherited text attributes. Hugging Face serialization restores the nested release schema; a decoder-prefix override preserves the explicit RoPE/window schedule.

_output_multiplier

= float(text['output_multiplier'])

_qk_scale_factor

= float(text['qk_scale_factor'])

layer_sliding_windows

layer_types

= text['layer_types'][:(self.num_hidden_layers)]

model_type

= 'muse_spark'

no_rope_layers

output_multiplier

float

Return the release’s output scaling factor.

router_output_multiplier

float

Return the matching MetaP router scaling factor.

scale_query_by

float

Return the release’s post-normalization query scale.

sliding_window

= int(text['sliding_window'])

vision_navit_config

dict[str, Any]

Translate the HF vision schema for the native NaViT implementation.

nemo_automodel.components.models.muse_spark.configuration_muse_spark.MuseSparkConfig.from_dict(
    config_dict: dict[str, typing.Any],
    kwargs: typing.Any = {}
) -> nemo_automodel.components.models.muse_spark.configuration_muse_spark.MuseSparkConfig | tuple[nemo_automodel.components.models.muse_spark.configuration_muse_spark.MuseSparkConfig, dict[str, typing.Any]]

classmethod

Apply prefix overrides before constructing the attention schedules.

nemo_automodel.components.models.muse_spark.configuration_muse_spark.MuseSparkConfig.layer_uses_rope(
    layer_idx: int
) -> bool

Return the release’s explicit per-layer RoPE choice.

nemo_automodel.components.models.muse_spark.configuration_muse_spark.MuseSparkConfig.to_dict() -> dict[str, typing.Any]

Serialize the HF nested schema, including native decoder-prefix overrides.