NVIDIA-NeMo/Automodel

🚀 NeMo AutoModel – What It Is

NeMo AutoModel (hosted under the NVIDIA‑NeMo organization) is a Python library that lets you fine‑tune or pre‑train huge language, vision‑language, and diffusion models with as little boilerplate as possible. It builds on top of PyTorch and NVIDIA’s parallel‑training stack (DTensor, custom kernels, tensor‑parallelism, pipeline‑parallelism, etc.) and provides ready‑made recipes for dozens of state‑of‑the‑art models that live on Hugging Face.

---

Core ideas

Idea	What it means for you
Model‑agnostic recipes	Each supported model (e.g., DeepSeek‑V4.1‑Flash, GLM‑5.3, Qwen‑3.8‑Flash‑Next, Nemotron‑3‑Ultra, Gemma‑4, etc.) has a YAML‑defined training script that configures data loading, parallelism, optimizer, and any special kernels. You just point the recipe at your data and launch it.
Full‑parameter and LoRA/PEFT support	You can train every weight of a model or use parameter‑efficient fine‑tuning (LoRA, PEFT) – the same interface works for both.
Speculative decoding & drafters	Built‑in support for training “draft” models (EAGLE, DFlash, DSpark) that enable faster inference via speculative decoding.
Agent‑friendly skills	A small collection of command‑line “skills” (e.g., run_recipe, model_onboard) that can be invoked by automation agents or notebooks.
NVIDIA‑optimized kernels	Recipes automatically pick custom kernels (e.g., CSA2, FlexAttention, TileLang, cuDNN DSA) to squeeze performance out of H100/GB200 GPUs.

---

What you can do with it

• Fine‑tune LLMs (dense, MoE, hybrid‑attention) on classic benchmarks like HellaSwag, MedPix, or custom datasets.
• Fine‑tune vision‑language models (LLaVA‑OneVision, Qwen‑VL, Inkling, MiniMax‑M3) with packed‑sequence or multi‑token prediction recipes.
• Train diffusion or D‑LLM models (DiffusionGemma, DFlash) and experiment with speculative decoding.
• Scale from a single GPU to dozens of H100/GB200 nodes using the same recipe; the library handles tensor‑ and pipeline‑parallelism under the hood.
• Add new models by writing a short model‑coverage MD file and a recipe – the repo already ships with a large catalog.

---

Getting started (quick‑start)

# 1. Install (Python 3.10+ required)
pip install nemo-automodel

# 2. Pick a recipe – e.g., fine‑tune DeepSeek‑V4.1‑Flash on HellaSwag
python -m nemo_automodel.run \
    --recipe examples/llm_finetune/deepseek_v41/deepseek_v41_flash_hellaswag_ep64_16nodes.yaml \
    --data_path /path/to/hellaswag

The YAML file contains sections for model, trainer, parallelism, and dataset. No code changes are needed unless you want to customize the data pipeline.

---

Documentation & resources

• Full docs: https://docs.nvidia.com/nemo/automodel/latest/index.html
• Model coverage list: https://docs.nvidia.com/nemo/automodel/latest/model-coverage/overview.html (shows every supported model and its capabilities)
• Performance summary: https://docs.nvidia.com/nemo/automodel/latest/performance-summary.html
• Examples folder: https://github.com/NVIDIA-NeMo/Automodel/tree/main/examples (recipes for LLM, VLM, diffusion, speculative decoding, etc.)
• Contributing guide: https://github.com/NVIDIA-NeMo/Automodel/blob/main/CONTRIBUTING.md

---

Who should look at this?

• Researchers who need a reliable, high‑performance training pipeline for the newest LLM/VLM architectures.
• Engineers building production fine‑tuning services on NVIDIA GPUs.
• ML‑ops teams that want a single repo to handle everything from data onboarding to multi‑node training.

---

License

The repository is released under the Apache 2.0 license (see the badge in the README).

---

TL;DR

NeMo AutoModel is NVIDIA’s turnkey library for fine‑tuning a massive catalog of modern language, vision‑language, and diffusion models at scale, with both full‑parameter and parameter‑efficient options, and with built‑in support for speculative decoding and NVIDIA‑optimized kernels.