Soup
A CLI that fine-tunes and post-trains LLMs from a single YAML config and one command, using layer streaming to train an 8B model on a 4 GB laptop GPU.
Soup turns LLM fine-tuning into a simple workflow: one config file, one command. Its layer-streaming feature keeps the frozen base model out of VRAM and feeds it to the GPU one decoder layer at a time, letting an 8B model such as Llama-3.1-8B-Instruct with NF4 quantization train on a 4 GB laptop GPU, measured at 119.6 tokens per second while remaining bit-exact against a normal resident run, and reproduced independently on an H100 in the same memory footprint.
The project states that even experienced teams spend 30-50% of their time fighting infrastructure rather than improving models, and aims to remove that by handling batch size, GPU detection and quantization automatically, training entirely on a user's own GPU with QLoRA with no cloud required and no SSH into a remote box.
Its latest release makes an unknown configuration key in its config file a hard failure instead of a silent warning, after a previous release found that six validated, documented training options were silently ignored by one of its backends.