oumi-ai/oumi

Oumi – an open‑source platform for the whole lifecycle of foundation models

What it is – Oumi is a Python‑based, fully open‑source toolkit that lets you prepare data, train, evaluate, and deploy large language models (LLMs) and multimodal models. It ships a command‑line interface (oumi) and a library that abstracts away the boilerplate of distributed training, LoRA/QLoRA fine‑tuning, data synthesis with LLM judges, and inference with engines such as vLLM or SGLang.

Key capabilities

• Training & fine‑tuning – supports models from 10 M to 405 B parameters, with SFT, LoRA, QLoRA, GRPO, and other state‑of‑the‑art methods. Distributed training back‑ends include FSDP, DeepSpeed and DDP.
• Multimodal support – works with text‑only and vision‑language models (Llama, Qwen, DeepSeek, Phi, InternVL, etc.).
• Data synthesis & curation – built‑in “LLM‑as‑a‑Judge” pipelines let you filter and generate training data automatically.
• Inference – one‑click deployment to popular inference servers (vLLM, SGLang) and cloud endpoints (Fireworks, Together, Parasail, etc.) via the oumi deploy CLI.
• Evaluation – a unified benchmark suite that can run standard LLM evals and vision‑language tests.
• Cloud‑native job management – oumi launch can submit jobs to AWS, Azure, GCP, Lambda, or other clusters, handling resource provisioning and monitoring.
• Recipes – a growing collection of ready‑made YAML configs for dozens of models (Qwen‑3, Llama 4, DeepSeek R1, etc.) covering training, inference, and evaluation.

Why you might use it

• Zero boilerplate – start a full training/eval/deploy pipeline in minutes by picking a recipe.
• Research‑ready – reproducible experiments, easy swapping of components, and support for cutting‑edge techniques.
• Enterprise‑grade – validated on large‑scale runs, with production‑level reliability and CI testing.
• Open & community‑driven – Apache‑2.0 licensed, pip‑installable, and backed by an active Discord and blog.

Getting started

# Install (CPU only)
uv pip install oumi
# Install with GPU support
uv pip install 'oumi[gpu]'

# Quick training example
oumi train -c configs/recipes/smollm/sft/135m/quickstart_train.yaml

# Evaluate the same model
oumi evaluate -c configs/recipes/smollm/evaluation/135m/quickstart_eval.yaml

# Deploy an inference endpoint
oumi deploy -c configs/recipes/smollm/inference/135m_infer.yaml

You can also run the CLI inside the official Docker image or use the one‑line installer script (curl … | bash). Detailed notebooks are provided for a tour, LoRA fine‑tuning, model distillation, evaluation, remote cloud training, and LLM‑as‑a‑Judge.

Documentation & community – Full docs at https://oumi.ai/docs, a blog with release announcements, a Discord server, and a public PyPI package (oumi).

---

All information above is taken directly from the repository’s README; no additional features have been inferred.