OpenViSTAI · Agents · Blockchain · Crypto
Submit a project
Menu
AI Developer Tools & SDKs

Soup

Soup is an open-source Python command-line tool for fine-tuning and post-training large language models using YAML configuration. It supports local workflows and documents training methods including SFT and DPO, with an optional layer-streaming mode intended to reduce GPU memory use.

Explore the practical guide ↓

Independent OpenViST project profile, based on linked sources.

Soup — OpenViST project profile in AI Developer Tools & SDKs.
OpenViST project illustration. See the linked sources for product details.
Understand the project

Practical overview

When it fits

Use Soup when you want a locally configured LLM training workflow and are comfortable working with Python, model assets and training data. Consider its beta layer-streaming mode only when its documented status and hardware-specific caveats fit your experiment.

Who it is for

  • Developers adapting language models with their own training data.
  • Practitioners who want a YAML-configured local training workflow.
  • Experimenters exploring LoRA, DPO or memory-constrained fine-tuning.

Documented uses

  • Define a model, task, dataset and training settings in a YAML configuration.
  • Run a configured fine-tuning or post-training workflow from the CLI.
  • Explore local GPU training, including a beta layer-streaming option for lower-VRAM setups.
Before you choose

Requirements & limits

The project describes Soup as free and Apache-2.0 licensed, with no paid tier. The README specifies Python 3.10–3.12.

  • The training extra is required for soup train; the base install is a lighter CLI.
  • The README recommends CUDA, supports Apple Silicon through MPS, and labels CPU training experimental and very slow.
  • For 7B QLoRA training, the requirements list 8 GB or more of VRAM.
  • Layer streaming is opt-in and beta; the cited 4 GB measurements predate a correctness repair and are pending remeasurement.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Install Soup and create a starter training configuration

Install the training-enabled CLI and generate a chat configuration to edit for a fine-tuning workflow.

Before you start

  • Python 3.10, 3.11 or 3.12.
  • A Python environment; the README recommends pipx or uv tool for an isolated CLI install.
  • The train extra is needed for the soup train command.

Steps

  1. Install with pipx install "soup-cli[train]" or, inside an existing environment, the command below.

    Documented command
    pip install "soup-cli[train]"

  2. Run soup init --template chat to create a starter configuration.

  3. Review and edit soup.yaml to specify the base model, training data, task and training options.

  4. Use soup train to start the configured fine-tuning workflow.

Expected outcome

A training-enabled Soup installation and a YAML-configured fine-tuning run.

Guide limitations

  • This recipe does not guarantee a successful run on a particular machine or model.
  • Soup lists CUDA as recommended, Apple Silicon MPS as supported, and CPU as experimental and very slow.
  • The README lists 8 GB or more of VRAM for 7B models with QLoRA; layer streaming is separately described as beta.
Documentation 1 ↗
Evidence and transparency

Sources & review

Official repository ↗
Source record details
Source availability
Active
Sources checked
2 October 2026
Original website title
MakazhanAlpamys/Soup · Official homepage title · retained in its source language
How we review sourcesSuggest a correction