Unsloth

Unsloth is an open-source library that makes fine-tuning language models much faster and lighter on memory. It rewrites key operations by hand so you can train larger models on a single consumer GPU, or in a free Colab notebook.

Ready-made notebooks cover popular models, and it publishes optimized quantized versions of new releases quickly. Reinforcement-learning training is supported as well.

It is beloved by hobbyists and startups who cannot afford big clusters.