Unsloth

Unsloth is an open-source library that makes fine-tuning open LLMs roughly 2x faster with far less memory via custom Triton kernels, enabling training of models like Llama, Qwen, and Gemma on modest GPUs.