NeMo Framework is NVIDIA's GPU accelerated, fully open-source, end-to-end training framework for large language models (LLMs), multi-modal models, diffusion and speech models. It enables seamless scaling of pretraining, post-training, and reinforcement learning workloads from single GPU to thousand-node clusters for both 🤗Hugging Face/PyTorch and Megatron models. This GitHub organization includes a suite of libraries and recipe collections to help users train models from end to end.
NeMo Framework is also a part of the NVIDIA NeMo software suite for managing the AI agent lifecycle.
| Installation | Checkpoint Conversion HF<>Megatron | LLM example recipes and scripts | VLM example recipes and scripts | |
|---|---|---|---|---|
| 1 ～ 1,000 GPUs | NeMo Automodel, NeMo RL | No Need | Pre-training, SFT, LoRA, DPO, GRPO | SFT, LoRA, GRPO |
| Over 1,000 GPUs | NeMo Megatron-Bridge, NeMo RL | Conversion | Pretrain, SFT, and LoRA, DPO with megatron_cfg, GRPO with megatron_cfg | SFT, LoRA, GRPO megatron config |
Visit the individual repos to find out more 🔍, raise 🐛, contribute ✍️ and participate in discussion forums 🗣️!
Note: The NeMo Framework is currently in the process of restructuring. The original NeMo 2.0 repository will now focus specifically on speech-related components, while other parts of the framework are being modularized into separate libraries such as NeMo Automodel, NeMo Gym, NeMo RL, and more. This transition aims to make NeMo more modular and developer-friendly.
| Repo | Key Functionality & Documentation Link | Training Loop | Training Backends | Inference Backends | Model Coverage | Container |
|---|---|---|---|---|---|---|
| NeMo Megatron-Bridge | Pretraining, LoRA, SFT | PyT native loop | Megatron-core | NA | LLM & VLM | NeMo Framework Container |
| NeMo AutoModel | Pretraining, LoRA, SFT | PyT native loop | PyTorch | NA | LLM, VLM, Omni, VFM | NeMo AutoModel Container |
| NeMo Speech | Pretraining,SFT | PyTorch Lightning Loop | Megatron-core & PyTorch | RIVA | Speech | NA |
| NeMo RL | SFT, RL | PyT native loop | Megatron-core & PyTorch | vLLM | LLM, VLM | NeMo RL container |
| NeMo Gym | RL Environment, Eval, Sandboxing | NA | NA | NA | NA | NeMo RL Container (WIP) |
| NeMo Curator | Data curation | NA | NA | NA | Agnostic | NeMo Curator Container |
| NeMo Guardrails | Guardrail model response | NA | NA | NA | NA | |
| NeMo Evaluator | Model evaluation | NA | NA | Agnostic | NeMo Framework Container | |
| NeMo Export-Deploy | Export to Production | NA | NA | vLLM, TRT, TRTLLM, ONNX | Agnostic | NeMo Framework Container |
| NeMo Run | Experiment launcher | NA | NA | NA | Agnostic | NeMo Framework Container |
| NeMo Skills | Reference pipeline for SDG & Eval | NA | NA | NA | Agnostic | NA |
| NeMo Emerging Optimizers | Collection of Optimizers | NA | Agnostic | NA | NA | NA |
| Nemotron | Developer asset hub for Nemotron models | NA | NA | NA | Nemotron models | NA |
| NeMo Data Designer | Synthetic data generation library | NA | NA | NA | NA | NA |
| NeMo Safe Synthesizer | Privacy-preserving synthetic tabular data | HF Trainer | PyTorch, Opacus | vLLM | LLM | NA |
| NeMo Anonymizer | PII detection and anonymization | NA | NA | NA | Agnostic | NA |
| Switchyard | Model routing, API translation & observability (OpenAI / Anthropic) | NA | NA | Agnostic (OpenAI-compatible) | Agnostic | NA |
The NeMo GitHub Org and its repo collections are created to address the following problems
- Need for composability: The Previous NeMo 2.0 version is monolithic and encompasses too many things, making it hard for users to find what they need. Container size is also an issue. Breaking down the Monolithic repo into a series of functional-focused repos to facilitate code discovery.
- Need for customizability: The Previous NeMo 2.0 version uses PyTorch Lighting as the default trainer loop, which provides some out of the box functionality but making it hard to customize. NeMo Megatron-Bridge, NeMo AutoModel, and NeMo RL have adopted pytorch native custom loop to improve flexibility and ease of use for developers.
Apache 2.0 licensed with third-party attributions documented in each repository.
