Soup

Soup is an open-source command-line tool that fine-tunes large language models from a single YAML configuration file. Published on GitHub with more than 9,400 stars, it targets a painful gap: most fine-tuning stacks assume a datacenter GPU, while Soup can train an 8B model on a laptop GPU with just 4 GB of memory.

Overview

The trick is layer streaming. Instead of loading the whole model into VRAM, Soup streams layers through the GPU as they are needed, drastically cutting memory while keeping training practical on consumer hardware. You describe the model, dataset, method and hyperparameters in one readable YAML file, run one command, and Soup handles data preparation, training and export. It wraps proven techniques from the Hugging Face ecosystem, including PEFT adapters and TRL trainers, but hides their boilerplate so a single file fully specifies a reproducible run.

Key Features

• One YAML file defines the entire fine-tuning run: model, data, method and hyperparameters
• Layer streaming trains models far larger than your VRAM, including 8B models on a 4 GB laptop GPU
• Multiple training methods: full SFT plus LoRA and QLoRA adapters, and preference methods such as DPO
• Built on PyTorch, Transformers, PEFT and TRL, with Hugging Face dataset support
• Export to GGUF for local inference with Ollama and llama.cpp workflows
• Sensible defaults and validation that catch configuration mistakes before training starts

💳 Pricing

Soup is free and open source under the Apache License 2.0. There is no paid tier; if you rent a GPU to run it, that rental is your only cost.

Who Is It For

Soup is for independent developers, students and small teams who want to fine-tune open models on hardware they already own. It removes the two usual blockers, cloud GPU budgets and framework complexity, making it an approachable on-ramp to customizing LLMs for a specific domain, style or language.