transformer_lens.model_bridge.supported_architectures.openai_gpt module¶
OpenAI GPT (GPT-1) architecture adapter.
The original GPT (OpenAIGPTLMHeadModel, openai-community/openai-gpt).
GPT-2’s block internals under different top-level names — combined-QKV
c_attn Conv1D, c_fc/c_proj MLP — but POST-norm (ln_1(x + attn),
ln_2(n + mlp)), no final LayerNorm, and embeddings at
transformer.tokens_embed / transformer.positions_embed.
- class transformer_lens.model_bridge.supported_architectures.openai_gpt.OpenAIGPTArchitectureAdapter(cfg: Any)¶
Bases:
ArchitectureAdapterArchitecture adapter for OpenAIGPTLMHeadModel (GPT-1) models.
- __init__(cfg: Any) None¶
Initialize the OpenAI GPT architecture adapter.