transformer_lens.model_bridge.supported_architectures.muse_glimmer module

Muse Glimmer (MuseGlimmerForConditionalGeneration) architecture adapter.

class transformer_lens.model_bridge.supported_architectures.muse_glimmer.MuseGlimmerArchitectureAdapter(cfg: Any)

Bases: ArchitectureAdapter

Architecture adapter for MuseGlimmerForConditionalGeneration models.

apply_output_logits_transform(logits: Tensor) → Tensor

Apply the architecture’s declared post-unembedding transform.

Most causal decoders only apply the shared optional Gemma-style soft cap. Adapters with additional scaling override this method instead of relying on similarly named configuration fields used by unrelated architectures.

component_mapping: ComponentMapping | None
supports_fold_ln = False
uses_split_attention: bool
weight_processing_conversions: dict