transformer_lens.model_bridge.supported_architectures.blenderbot module

Blenderbot (BlenderbotForConditionalGeneration) adapter: Pegasus-style pre-norm encoder-decoder with learned pos-embeds and baked-in embedding scale (so embed hooks see already-scaled output); asymmetric stacks, n_layers follows the decoder.

class transformer_lens.model_bridge.supported_architectures.blenderbot.BlenderbotArchitectureAdapter(cfg: Any)

Bases: BartFamilyArchitectureAdapter

Architecture adapter for BlenderbotForConditionalGeneration models.

component_mapping: ComponentMapping | None
has_final_stack_norm: bool = True
n_layers_from: str = 'decoder'
require_symmetric_layers: bool = False
uses_split_attention: bool
weight_processing_conversions: Dict[str, ParamProcessingConversion | str] | None