Unsloth

Unsloth makes fine-tuning open LLMs practical on consumer hardware: hand-written Triton kernels train LoRA/QLoRA adapters roughly 2x faster while using far less VRAM than stock Hugging Face training, so a 7B-class model can be fine-tuned on a single gaming GPU. The pip library covers SFT, DPO/ORPO, GRPO, and vision models; the newer Unsloth Desktop app (Windows/macOS/Linux) offers a native GUI for running and training models without touching the terminal. 77.9k GitHub stars, Apache-2.0. Treat the '2x faster' figures as vendor benchmarks — re-test on your own workload.

```bash

pip install unsloth

```

```python

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained("unsloth/Llama-3.2-3B-Instruct")

model = FastLanguageModel.get_peft_model(model, r=16, target_modules=["q_proj","v_proj"])

# then train with TRL's SFTTrainer as usual

```

Or download Unsloth Desktop for your OS from the GitHub releases page. Docs: unsloth.ai/docs.