Soup CLI enables preference tuning on 4GB GPUs via layer streaming
August 4, 2026
Soup-cli automates LLM fine-tuning using a single YAML configuration and layer streaming to minimize VRAM usage. This method keeps the frozen base model out of memory, allowing DPO, ORPO, SimPO, and KTO to run on consumer hardware like 4GB laptop GPUs.
HOW THIS AFFECTS YOU
●
builderYou can perform complex preference alignment on local consumer hardware without managing complex infra or SSH.
●
researcherYou can test new alignment loss functions like SimPO or KTO on low-memory systems using layer streaming.