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.