Unsloth

What it is
Unsloth is an open-source library designed to make QLoRA fine-tuning of LLaMA models 2x faster and memory-efficient. Ideal for running on Colab, Kaggle, or low-VRAM machines.

Key features

• Up to 2x speedup: Optimized CUDA kernels and attention rewrites.
• Low RAM/GPU support: Fine-tune LLaMA 2/3 on 8GB or even 4GB GPUs.
• QLoRA optimized: Flash-attn, 4-bit quantization, paged attention built-in.
• Trainer integration: Works with Hugging Face transformers and trl.
• Colab/Kaggle-ready: Official notebooks and support for budget training.