Soup CLI is an innovative, open-source developer tool designed to democratize LLM fine-tuning by enabling developers to train large language models on modest consumer hardware. Specifically, Soup CLI makes it possible to fine-tune an 8B LLM using just a 4 GB laptop GPU.

Because Low-Rank Adaptation (LoRA) keeps the base model frozen—meaning it is read from and never written to—Soup CLI keeps the base model in system RAM and streams it into the GPU one decoder layer at a time. This reduces peak VRAM consumption to just a single layer instead of holding the entire model in memory. Tested on an RTX 3050 Laptop 4 GB GPU, Llama-3.1-8B trains at an impressive 119.6 tokens/sec while maintaining a peak usage of only 3.32 GB.

Featuring a streamlined workflow requiring just one YAML configuration file and a single command, Soup CLI supports multiple training methods including SFT, DPO, GRPO, and KTO, alongside built-in capabilities for evaluation, gating, and model export. Distributed under the Apache-2.0 license, the project prides itself on radical transparency, publishing all performance numbers—including experimental data that didn't make the cut.

Pros of Soup CLI

Fine-tune an 8B LLM (like Llama-3.1-8B) on a modest 4 GB laptop GPU
Extremely memory efficient by streaming the frozen base model layer-by-layer from system RAM
Simple user experience with a single YAML configuration file and one-command execution
Supports multiple training methods including SFT, DPO, GRPO, and KTO, plus eval, gating, and export
Open-source (Apache-2.0) with transparent performance data and published metrics

Cons of Soup CLI

Requires sufficient system RAM to hold the base model while streaming to the GPU
Performance heavily relies on system memory bandwidth for layer streaming
Limited to hardware configurations compatible with the CLI's memory-streaming architecture

Frequently Asked Questions

What is Soup CLI?

Soup CLI is an open-source developer tool that allows users to fine-tune large language models (such as an 8B LLM) on a low-end 4 GB laptop GPU by leveraging a layer-by-layer memory streaming approach.

How does Soup CLI reduce VRAM usage?

Because LoRA keeps the base model frozen and read-only, Soup CLI stores the base model in system RAM and streams it into the GPU one decoder layer at a time. This keeps peak VRAM consumption down to just a single layer rather than the whole model.

What training methods and features are supported?

Soup CLI supports SFT (Supervised Fine-Tuning), DPO, GRPO, and KTO. It also includes built-in evaluation, model gating, and model export functionalities, all configured via a single YAML file and executed with one command.