Soup brings LLM fine-tuning to low-memory laptop GPUs

Soup simplifies LLM fine-tuning and post-training through a YAML configuration and command-line workflow, with automatic GPU detection, batching and quantization. Its experimental layer-streaming feature enables training an 8B model on a 4 GB laptop GPU by keeping the frozen base model outside GPU memory and transferring decoder layers as needed. The feature remains opt-in and beta.

The v0.76.0 release addresses configuration settings that were accepted but never applied. Some now prevent configurations from loading, while others generate warnings ahead of future rejection. Validation splits now receive evaluation on supported training paths rather than merely being excluded from training. New tools check adapters for silent failures, clean datasets, verify training recipes and identify dataset rows associated with loss spikes.

Soup also provides a local browser dashboard, model chat, evaluation, adapter merging and deployment exports. Layer-streaming development has included a repair for incorrect gradients that could occur even when forward outputs matched and loss curves appeared healthy. Comparisons with resident training found indistinguishable trained-model quality, but published performance results span different software versions and hardware clock conditions and are not directly comparable. Soup is free, openly developed and licensed under Apache-2.0.