Soup: Fine-Tune LLMs from a Single YAML File

What is Soup?

Soup is an open-source toolkit that turns LLM fine-tuning into a one-command workflow. You describe the whole training run — model, data, hyperparameters, LoRA adapters, quantization — in a single YAML file, then run soup train. No SSH sessions, no brittle shell scripts, no config sprawl.

Layer Streaming: The Killer Feature

Fine-tuning a large model normally requires the full base model to fit in VRAM. Soup's layer-streaming engine instead feeds the GPU one frozen decoder layer at a time. Measured on an RTX 3050 Laptop (4 GB), Llama-3.1-8B-Instruct fine-tunes at 119.6 tokens/s with a 3.32 GB peak — bit-exact against a regular resident run, and independently reproduced on an H100 at 113.00 tok/s in the same 3.32 GB footprint.

Everything in One YAML

Model registry, adapters, quantization (NF4/FP8), datasets, and evaluation are all declared in soup.yaml. Configs can be loaded from local files or remote URLs, so a team can share one canonical training config.

Key Features

• Single YAML config — model, data, hyperparameters, everything
• Layer streaming — train 8B models on 4 GB laptop GPUs
• Opt-in stacks — light CLI, train, serve, UI, and data tools
• Chat UI & model registry — inspect and serve fine-tuned models
• Apache-2.0 — fully open source, self-hosted

Soup is ideal for indie hackers, students, and small teams who want full control over fine-tuning without a GPU cluster budget.