This paper presents PassLLM, a novel framework that adapts large language models (LLMs) for password guessing across multiple attack scenarios. By leveraging low-rank adaptation and customized generation algorithms, PassLLM enables efficient and effective password generation at scale. It supports both trawling and targeted guessing, including scenarios involving personal information and password reuse. The paper demonstrates that PassLLM consistently outperforms existing approaches in multiple real-world datasets. To support reproducibility and enable future research, we release the source code used in our study, including model fine-tuning, password generation, Monte Carlo evaluation, dynamic beam search, and model distillation. In particular, we also provide two trained model checkpoints that can reproduce some key experimental results in our paper (e.g., Figs. 9(a) and 9(b)).