@article{harlisloawarecolocationofllminferenceand,
  title = {Harli: SLO-Aware Co-location of LLM Inference and PEFT-based Finetuning on Model-as-a-Service Platforms},
  author = {Ao Xu and Han Zhao and Weihao Cui and Quan Chen and Yukang Chen and Shulai Zhang and Shuang Chen and Jiemin Jiang and Zhibin Yu and Minyi Guo},
  year = {2025},
  eprint = {2511.11729},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2511.11729},
}