`bridge.models.muse_glimmer.muse_glimmer_config`#

Native HybridModel configuration objects for Muse Glimmer.

Module Contents#

Classes#

`MuseGlimmerTransformerConfig`	Muse-specific decoder fields not represented by MCore’s base config.
`MuseGlimmerVisionModelConfig`	Serializable configuration for the replicated Muse vision encoder.
`MuseGlimmerModelConfig`	Complete builder configuration for Muse Glimmer.

Data#

__all__

API#

class bridge.models.muse_glimmer.muse_glimmer_config.MuseGlimmerTransformerConfig#

Bases: megatron.bridge.models.transformer_config.TransformerConfig

Muse-specific decoder fields not represented by MCore’s base config.

post_norm_epsilon: float#: 1e-08

output_multiplier: float#: 0.19611613513818404

final_logit_softcapping: float#: 20.0

class bridge.models.muse_glimmer.muse_glimmer_config.MuseGlimmerVisionModelConfig#

Serializable configuration for the replicated Muse vision encoder.

hidden_size: int#: 1536

intermediate_size: int#: 8960

num_hidden_layers: int#: 50

num_attention_heads: int#: 16

patch_size: int#: 14

patch_temporal: int#: 2

merge_size: int#: 2

pos_emb_height: int#: 32

pos_emb_width: int#: 32

max_position_embeddings: int#: 1024

layer_norm_epsilon: float#: 1e-05

hidden_activation: str#: ‘gelu’

rotary_base: float#: 10000.0

layer_types: list[str]#: ‘field(…)’

property depth: int#: Expose the vision depth expected by Bridge’s VLM FLOPs estimator.

property spatial_merge_size: int#: Expose the spatial merge factor expected by VLM data utilities.

class bridge.models.muse_glimmer.muse_glimmer_config.MuseGlimmerModelConfig#

Bases: megatron.bridge.models.common.ModelConfigOverrideMixin, megatron.training.models.hybrid.HybridModelConfig

Complete builder configuration for Muse Glimmer.

builder: ClassVar[str]#: ‘megatron.bridge.models.muse_glimmer.MuseGlimmerModelBuilder’

transformer_config_class: ClassVar[type[megatron.bridge.models.transformer_config.TransformerConfig]]#: None

hybrid_attention_layers_include_mlp: ClassVar[bool]#: True

vision: bridge.models.muse_glimmer.muse_glimmer_config.MuseGlimmerVisionModelConfig#: ‘field(…)’

image_token_id: int#: 200092

video_token_id: int#: 200091

bos_token_id: int | None#: 200000

eos_token_id: int | list[int] | None#: 200001

pad_token_id: int | None#: None

vision_output_size: int#: 6144

projector_hidden_size: int#: 4096

projector_hidden_activation: str#: ‘gelu’

freeze_language_model: bool#: False

freeze_vision_model: bool#: False

freeze_vision_projection: bool#: False

recompute_vision_layers: bool#: False

property vision_config: bridge.models.muse_glimmer.muse_glimmer_config.MuseGlimmerVisionModelConfig#: Expose the vision config to shared VLM training utilities.

property special_token_ids: dict[str, int]#: Return media token IDs used by multimodal data pipelines.

get_builder_cls() → type#: Resolve the Muse builder through Bridge’s target allowlist.

as_dict() → dict[str, Any]#: Serialize the Hybrid config with a symbolic activation function.

classmethod from_dict( data: dict[str, Any], ) → bridge.models.muse_glimmer.muse_glimmer_config.MuseGlimmerModelConfig#: Deserialize a Muse Hybrid config and restore its activation callable.

bridge.models.muse_glimmer.muse_glimmer_config.__all__#: [‘MuseGlimmerModelConfig’, ‘MuseGlimmerTransformerConfig’, ‘MuseGlimmerVisionModelConfig’]

bridge.models.muse_glimmer.muse_glimmer_config#

Module Contents#

Classes#

Data#

API#

`bridge.models.muse_glimmer.muse_glimmer_config`#