transformer_lens.model_bridge.supported_architectures.exaone4 module

EXAONE 4.0 architecture adapter.

LG AI Research’s EXAONE-4.0 (Exaone4ForCausalLM, native transformers — distinct from the remote-code EXAONE-3.x family in exaone.py): the OLMo-2 post-norm block shape (per-head Q/K RMSNorms, post_attention_layernorm / post_feedforward_layernorm inside the residual branch) plus hybrid sliding/global attention (layer_types) with global-NoPE gating.

class transformer_lens.model_bridge.supported_architectures.exaone4.Exaone4ArchitectureAdapter(cfg: Any)

Bases: Olmo2ArchitectureAdapter

Architecture adapter for Exaone4ForCausalLM models.

__init__(cfg: Any) → None

Initialize the EXAONE 4.0 architecture adapter.