Soup

Soup is an open-source LLM fine-tuning toolkit that lets you configure an entire training run from a single YAML file. Its signature feature is layer streaming: instead of loading the frozen base model into VRAM, Soup streams one decoder layer at a time through the GPU, so you can fine-tune an 8B-parameter model on a laptop with just 4 GB of VRAM. Measured on an RTX 3050 Laptop, Llama-3.1-8B-Instruct trains at 119.6 tokens per second with a 3.32 GB peak, bit-exact against a normal resident run. Soup ships with a CLI, a chat UI, data tools, a model registry, benchmarks, and templates, and supports loading configs from local files or remote URLs. Everything is opt-in: install the light CLI alone or add the train, serve, ui, and data stacks. Licensed under Apache-2.0.