parameter-efficient tuning aims to mitigate the large memory requirements of
adapting pretrained language models for downstream tasks. For example, one
popular method, prefix-tuning, prepends trainable tokens to sequences while
freezing the rest of the model's parameters. Although such