Skip to content
GitHub·

🤖Soup Simplifies LLM Training With One Command

Training LLMs just got a whole lot easier

TL;DR

Soup simplifies the process of fine-tuning and post-training LLMs, requiring no SSH or complex configurations. Install with 'pip install soup-cli[train]' and start training locally on your GPU.

Soup just made life easier for anyone working with large language models (LLMs). The tool streamlines the often painful process of fine-tuning and post-training these models into a single command, eliminating the need for SSH or complex configurations. If you're tired of wrestling with batch sizes and GPU detection, Soup's auto-everything approach is a game-changer. With version 0.72.4, Soup introduces layer streaming for DPO, ORPO, SimPO, and KTO, reducing VRAM usage by feeding the frozen base to the GPU one decoder layer at a time. This means you can fine-tune models locally on your own hardware with QLoRA, making it accessible even if cloud resources are limited.

Soup Simplifies LLM Training With One Command — GitHub

Key Points

1

pip install 'soup-cli[train]' installs the full stack for fine-tuning and post-training LLMs

2

Version 0.72.4 adds layer streaming for DPO, ORPO, SimPO, KTO; VRAM usage reduced by feeding layers one at a time

3

Layer streaming on RTX 3050 4GB peaked at 0.914× the supervised-fine-tuning peak with streamed DPO

4

Templates include chat, code, tool-calling, medical, reasoning, vision, kto, orpo, simpo, ipo, bco, rlhf, pretrain, moe, longcontext, embedding, audio

5

Training, testing, and shipping are handled by 'soup train --config soup.yaml'

Why It Matters

If you're training LLMs locally on your GPU with QLoRA, Soup's layer streaming feature reduces VRAM usage significantly. This means smaller models can now be trained without the need for cloud resources or high-end GPUs. However, the $0.20/GB-month premium only makes sense above ~100K read IOPS per instance, so smaller databases should stick with their current setup.

LLMtraininglocal-gpustreaming

Frequently Asked Questions

Why does this matter?

If you're training LLMs locally on your GPU with QLoRA, Soup's layer streaming feature reduces VRAM usage significantly. This means smaller models can now be trained without the need for cloud resources or high-end GPUs. However, the $0.20/GB-month premium only makes sense above ~100K read IOPS per instance, so smaller databases should stick with their current setup.

What happened?

Soup simplifies the process of fine-tuning and post-training LLMs, requiring no SSH or complex configurations. Install with 'pip install soup-cli[train]' and start training locally on your GPU.

Comments

Subscribe to join the conversation...

Be the first to comment

Enjoyed this article?

Get it daily. 7am. Free. Reads in 5 minutes.

Also get