transformer_lens.model_bridge.supported_architectures.bart module

BART adapter and the shared BART-family encoder-decoder base (BART, Marian, MBart, Pegasus, Blenderbot, M2M100/NLLB); per-member differences are declarative.

class transformer_lens.model_bridge.supported_architectures.bart.BartArchitectureAdapter(cfg: Any)

Bases: BartFamilyArchitectureAdapter

Architecture adapter for BartForConditionalGeneration models.

Post-LN with layernorm_embedding; checkpoints ship scale_embedding=False, so the family default-on is disabled.

component_mapping: ComponentMapping | None

force_scale_embedding: bool = False

has_layernorm_embedding: bool = True

uses_split_attention: bool

weight_processing_conversions: Dict[str, ParamProcessingConversion | str] | None

class transformer_lens.model_bridge.supported_architectures.bart.BartFamilyArchitectureAdapter(cfg: Any)

Bases: ArchitectureAdapter

Shared base for the BART-family encoder-decoder adapters.

__init__(cfg: Any) → None

Validate the config, set family flags, and build the mapping.

force_scale_embedding: bool = True

has_final_stack_norm: bool = False

has_layernorm_embedding: bool = False

n_layers_from: str = 'encoder'

require_symmetric_layers: bool = True

setup_hook_compatibility(bridge: Any) → None

Fold the trained final_logits_bias into the unembed bias.

HF adds the buffer after lm_head, so b_U would read fabricated zeros and unembed.hook_out would fire pre-bias (Marian opus-mt trains it). Moving it into the bias UnembeddingBridge injects is numerically identity; zeroing the buffer keeps the (re-run) fold idempotent.