soup

Soup is a command-line application designed to streamline the fine-tuning and post-training of large language models (LLMs). It aims to eliminate the complexity of infrastructure setup, allowing users to train models with a single YAML configuration file and one command. A key feature is layer streaming, which enables fine-tuning an 8B model on a 4 GB laptop GPU by streaming the frozen base model from host RAM one decoder layer at a time, achieving high throughput with low VRAM usage. The tool supports various training tasks (SFT, DPO, GRPO, PPO, KTO, ORPO, SimPO, IPO, BCO, etc.), multiple backends (transformers, MLX, Unsloth), and a wide range of models from the HuggingFace Hub. It includes a web UI for experiment management, data tools, evaluation, serving, and export capabilities. The project is open-source under Apache-2.0 and emphasizes measured performance over claims, with a focus on community contributions and hardware-gated development.