An open-source tool called "Soup" has been released, enabling fine-tuning of large-scale language models on GPUs found in typical laptops. According to the developers' testing, they fine-tuned the 8-billion-parameter "Llama 3.1 8B Instruct" using a GPU with 4GB of VRAM. Soup employs a technique called "layer streaming," which places decoder layers that do not change during training in main memory and sends the decoder layers required for computation one at a time to VRAM. The developers reportedly succeeded in quantizing Meta's "Llama 3.1 8B Instruct" in NF4 format and fine-tuning it with LoRA using a GeForce RTX 3050 Laptop GPU with 4GB of VRAM. Under conditions of processing a text length of 512 tokens at a time and a batch size of 1, peak VRAM usage was 3.32GB and processing speed was 119.6 tokens per second. They also confirmed calculation results matched bit by bit when the entire NF4 model was stored on the GPU. Running on a GPU with limited VRAM capacity uses CPU main memory in conjunction with GPU VRAM, resulting in longer training times. It won't work under all settings, and VRAM may become insufficient depending on text length and batch size. Soup is a command-line tool where fine-tuning starts by specifying training data and executing a command. Version 0.72 update added support for