Oumi is an open-source platform designed for the end-to-end lifecycle of foundation models, from data preparation and training to evaluation and deployment. It targets researchers and developers needing a flexible, production-grade solution for building and managing large language and vision-language models, offering a unified API across various hardware and cloud environments.

How It Works

Oumi provides a comprehensive toolkit that abstracts away complex distributed training and inference setups. It supports state-of-the-art techniques like SFT, LoRA, QLoRA, and DPO, and integrates with optimized inference engines such as vLLM and SGLang for efficient deployment. The platform also includes LLM-as-a-Judge for data curation and a robust evaluation framework, enabling users to manage model development from a laptop to large-scale clusters.

Quick Start & Requirements

• Install: pip install oumi (CPU/NPU) or pip install oumi[gpu] (GPU). Install from source for the latest version: pip install git+https://github.com/oumi-ai/oumi.git.
• Prerequisites: GPU support requires an Nvidia or AMD GPU.
• Documentation: oumi.ai, Getting Started, Model Finetuning Guide.

Highlighted Details

• Supports training and inference for models ranging from 10M to 405B parameters.
• Comprehensive support for numerous LLM and VLM architectures including Llama, Qwen, DeepSeek, Phi, and LLaVA.
• Features "Zero Boilerplate" with ready-to-use recipes for popular models and workflows.
• Enables remote training and job management on cloud platforms (AWS, Azure, GCP, Lambda).

Maintenance & Community

Oumi is a community-first project with an active Discord community. Contributions are welcomed via CONTRIBUTING.md.

Licensing & Compatibility

Licensed under Apache License 2.0, which is permissive for commercial use and closed-source linking.

Limitations & Caveats

Oumi is currently in beta and under active development, with some advanced features subject to change.