LMFlow

Extensible toolkit for finetuning and inference of large foundation models (8.5K★, NAACL 2024 demo track best paper), launched in the post-ChatGPT wave of 2023 and maintained through 2026: full/LoRA finetuning, RLHF, custom optimizers, and efficient inference in one framework. The same OptimalScale group produced LISA (layerwise importance sampling, arXiv 2403.17919), a memory-efficient finetuning method.

Built by Tong Zhang's group at HKUST; legacy anchor for the university's training-infrastructure line. Personnel note: Tong Zhang has since moved to UIUC.