File size: 977 Bytes
896bea2 b61a50c 896bea2 b61a50c | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 | ---
title: README
emoji: 📚
colorFrom: gray
colorTo: pink
sdk: static
pinned: false
license: apache-2.0
short_description: Fine-tune large models on small GPUs
---
# Soup
Open-source engine for fine-tuning large models on small GPUs.
**8B model fine-tuned in 3.3 GB of VRAM** — on a 4 GB RTX 3050 laptop GPU.
Against DeepSpeed ZeRO-3 on the same machine: **2.93x the speed, 9.7x less memory.**
Apache-2.0. Every measurement is published, including the runs that failed
and one explanation we posted and then retracted after measuring it.
- Repository: https://github.com/MakazhanAlpamys/Soup
- Benchmarks: https://github.com/MakazhanAlpamys/Soup/tree/main/benchmarks
- Paper: https://trysoup.dev/docs/paper
- Run it yourself on a free Colab T4: https://trysoup.dev/docs/layer-streaming#run-it-yourself-on-a-free-colab-t4
Install:
pip install "soup-cli[train]"
Ollama made running models something you do on your own machine.
We are doing that for training them. |