transformer_lens.model_bridge.supported_architectures.mbart module

MBart (MBartForConditionalGeneration) adapter: Bart layout but PRE-norm with an extra final LayerNorm per stack; embedding scale is baked in (embed hooks see it).

class transformer_lens.model_bridge.supported_architectures.mbart.MBartArchitectureAdapter(cfg: Any)

Bases: BartFamilyArchitectureAdapter

Architecture adapter for MBartForConditionalGeneration models.

component_mapping: ComponentMapping | None
has_final_stack_norm: bool = True
has_layernorm_embedding: bool = True
uses_split_attention: bool
weight_processing_conversions: Dict[str, ParamProcessingConversion | str] | None