Soup (soup-cli) is a Python CLI that fine-tunes and post-trains LLMs from one YAML config, supporting QLoRA, full fine-tuning, and preference losses (DPO, ORPO, SimPO, KTO).

Its layer streaming method feeds frozen base weights one decoder layer at a time into VRAM, training Llama-3.1-8B-Instruct at NF4 in 3.32 GB peak on a 4 GB RTX 3050 laptop GPU. The tool runs locally on Python 3.10–3.12 and auto-handles batch size, GPU detection, and quantization parameters.