Soup
🎁FREEGitHub
Soup is an open-source CLI tool that lets you fine-tune and post-train Large Language Models with a single YAML config and one command. Its standout feature is layer streaming, which makes it possible to train an 8B model on a 4 GB laptop GPU.
✨Key Features
- ▸One-command workflow: soup init → edit a simple YAML → soup train.
- ▸Layer Streaming (BETA): Streams the frozen base model layer-by-layer so you can fine-tune 8B models on very low VRAM (measured ~3.3 GB peak on RTX 3050 4 GB).
- ▸Full post-training suite: Supports SFT, DPO, GRPO, PPO, KTO, ORPO, SimPO, reward modeling, distillation, vision, tool-calling, and more.
- ▸Huge model compatibility: Works with virtually any Hugging Face causal LM (Llama, Qwen, Gemma, DeepSeek, etc.).
- ▸Local-first & auto-magic: Automatic batch size, quantization (QLoRA/NF4), GPU detection. No complex multi-GPU or cloud setup required.
- ▸Extra capabilities: Serving (OpenAI-compatible + Anthropic Messages), evaluation, data tools, Docker support, Web UI.
- ▸Actively developed: ~5.7k GitHub stars, frequent releases (currently around v0.74), Apache-2.0 license.
Advertisement
728 x 90 Ad Space