transformer_lens.model_bridge.supported_architectures.m2m100 module

M2M100 architecture adapter.

Covers Meta’s M2M100 and NLLB-200 translation families (M2M100ForConditionalGeneration). Same q/k/v/out_proj + fc1/fc2 layout as Bart/Marian, but PRE-norm (LayerNorm before attention and MLP) with an extra final LayerNorm after each stack, deterministic sinusoidal positions with a padding-aware offset, and the sqrt(d_model) embedding scale baked into M2M100ScaledWordEmbedding itself — so unlike Marian, hooks on embed observe the already-scaled output.

class transformer_lens.model_bridge.supported_architectures.m2m100.M2M100ArchitectureAdapter(cfg: Any)

Bases: BartFamilyArchitectureAdapter

Architecture adapter for M2M100ForConditionalGeneration (M2M100 / NLLB) models.

component_mapping: ComponentMapping | None

has_final_stack_norm: bool = True

uses_split_attention: bool

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