Great point. Although in the article I didn't explore the hypermeters of PromptEncoderConfig, it is the total number of virtual tokens of the prompt encoder. The source code is https://github.com/huggingface/peft/blob/main/src/peft/tuners/p_tuning.py#L32