Soup

What it does

Soup is a CLI for fine-tuning and post-training LLMs with one config and one command. Layer streaming feeds the frozen base model to the GPU one layer at a time.

Why it is useful

• The post reports an 8B model training in about 3.3 GB of peak VRAM.
• No SSH into a remote GPU box needed.

How to use it

pip install "soup-cli[train]"

Write the YAML config, then run the training command from the README.

Watch out for

• Streaming layers is slower than keeping the whole model on the GPU.