PyTorch-native open-source library for fine-tuning LLMs — LoRA, QLoRA, full fine-tune, RLHF — with memory-efficient recipes for consumer GPUs. Fully hackable Python.