Auto-matching hidden layers in Pytorch LLMs — LessWrong