Important

You are viewing the NeMo 2.0 documentation. This release introduces significant changes to the API and a new library, NeMo Run. We are currently porting all features from NeMo 1.0 to 2.0. For documentation on previous versions or features not yet available in 2.0, please refer to the NeMo 24.07 documentation.

NeMo AutoModel#

NeMo AutoModel is a worfkow within the NeMo machine learning framework that supports both pretraining and fine-tuning Large Language Models (LLMs), designed to accelerate your journey from data to deployment. Whether you’re a researcher or a developer, NeMo AutoModel makes it incredibly easy to fine-tune state-of-the-art models from Hugging Face with minimal setup. You’ll benefit from:

Day-0 Support: Models available on Hugging Face Hub can be fine-tuned instantly on your own custom dataset.
Seamless Hugging Face Integration: Models fine-tuned with NeMo AutoModel are fully compatible with the Hugging Face ecosystem.
Rapid Experimentation: Start experimenting in minutes using ready-made scripts, notebooks, and run recipes.
Distributed Fine-Tuning: Scale your fine-tuning with Fully Sharded Data Parallelism 2 (FSDP2).
Accelerated Iteration: Enables rapid experimentation by streamlining model fine-tuning, evaluation, and deployment, so you can iterate faster on your research and applications.

Quickstart#

Ready to start? Here are the best resources to dive in:

For Parameter-Efficient Fine-Tuning (PEFT):
- Read the PEFT User Guide automodel/peft.rst.
- Try the Quickstart using the standalone python3 PEFT script.
- Follow the guided Jupyter Notebook.
- Access the NeMo-Run recipe.
For Supervised Fine-Tuning (SFT):
- Consult the SFT User Guide automodel/sft.rst.
- Run the standalone python3 SFT script.
- Explore the Jupyter Notebook.
- Check out the Multinode Jupyter Notebook.
- See the NeMo-Run recipe.

Start exploring today and unlock the power of your LLMs!

Known Issues#

NeMo AutoModel is currently in preview release; limitations are expected in performance and memory usage.