transformer_lens.model_bridge.supported_architectures.nanochat module

NanoChat (NanoChatForCausalLM) adapter: Llama-style decoder with weightless RMSNorm (so nothing for fold_ln to fold), ungated relu^2 MLP, and soft-capped logits; attention delegated (full-width q/k norm after RoPE).

class transformer_lens.model_bridge.supported_architectures.nanochat.NanoChatArchitectureAdapter(cfg: Any)

Bases: ArchitectureAdapter

Architecture adapter for NanoChatForCausalLM models.

__init__(cfg: Any) None

Initialize the NanoChat architecture adapter.

component_mapping: ComponentMapping | None
supports_fold_ln = False
uses_split_attention: bool
weight_processing_conversions: Dict[str, ParamProcessingConversion | str] | None