Soup CLI is a command-line tool for fine-tuning Large Language Models (LLMs). It enables training models that are larger than available VRAM by streaming layers from CPU RAM or NVMe. Soup CLI supports various fine-tuning techniques like LoRA, DPO, and QLoRA, and offers features for data preparation, model evaluation, and deployment.
Free
How to use Soup CLI?
Install via pip: `pip install soup-cli`. Use commands like `soup train` for fine-tuning, `soup draft measure` to evaluate speculative decoding, and `soup ship` for model deployment. The CLI offers extensive options for configuring training parameters, data handling, and model optimization, making LLM fine-tuning accessible on consumer hardware.
Soup CLI 's Core Features
Layer streaming for training LLMs larger than VRAM
Support for LoRA, QLoRA, DPO, and other fine-tuning methods
Speculative decoding measurement and distillation
Automated model evaluation and governance features
Data preparation and augmentation tools
Efficient inference serving with model registry
Cross-tokenizer support for draft models
Soup CLI 's Use Cases
Fine-tuning large LLMs on consumer GPUs
Training models that exceed available VRAM
Evaluating and optimizing speculative decoding performance