Soup: Fine-tune LLMs from One YAML, Train 8B Model on 4GB GPU

MakazhanAlpamys · github · 2026-08-15

Soup is a Python tool that fine-tunes LLMs from a single YAML config. It uses layer streaming to train an 8B model on a 4GB consumer GPU. Supports LoRA, QLoRA, SFT, DPO, and integrates with GGUF, Ollama. Gained 303 stars today, total 1,431.

Original post →

More from coding & agent

coding & agent channel →