NVIDIA NeMo™ is an open-source framework for developers to build and train state-of-the-art conversational AI models.
What Is NVIDIA NeMo?
NVIDIA NeMo, part of the NVIDIA AI platform, is a framework for building, training, and fine-tuning GPU-accelerated speech and natural language understanding (NLU) models with a simple Python interface. Using NeMo, developers can create new model architectures and train them using mixed- precision compute on Tensor Cores in NVIDIA GPUs through easy-to-use application programming interfaces (APIs).
NeMo Megatron is a part of the framework that provides parallelization technologies such as pipeline and tensor parallelism from the Megatron-LM research project for training large-scale language models.
With NeMo, you can build models for real-time automated speech recognition (ASR), natural language processing (NLP), and text-to-speech (TTS) applications such as video call transcriptions, intelligent video assistants, and automated call center support across healthcare, finance, retail, and telecommunications.
Rapid Model Building
Configure, build, and train models quickly with simple Python APIs.
Download and customize pre-trained state-of-the-art models from the NVIDIA NGC™ catalog.
Interoperate NeMo with PyTorch and the PyTorch Lightning ecosystem.
Deploy with NVIDIA AI
Leverage NVIDIA AI by applying NVIDIA® TensorRT™ optimizations and exporting to NVIDIA Riva for high performance inference.
Ping An addresses millions of queries from customers each day using chatbot agents. As an early partner of the Riva early access program, we were able to use the tools and build better solutions with higher accuracy and lower latency, thus providing better services. More specifically, with NeMo, the pre-trained model, and the ASR pipeline optimized using Riva, the system achieved 5 percent improvement on accuracy, so as to serve our customers with better experience.
— Dr. Jing Xiao, Chief Scientist at Ping An
In our evaluation of Riva for virtual assistants and speech analytics, we saw remarkable accuracy by fine-tuning the automated speech recognition models in the Russian language using the NeMo toolkit in Riva. Riva can provide up to 10X throughput performance with powerful TensorRT optimizations on models, so we’re looking forward to using Riva to get the most out of these technology advancements.
— Nikita Semenov, Head of Machine Learning at MTS AI
InstaDeep delivers decision-making AI products and solutions for enterprises. For this project, our goal is to build a virtual assistant in the Arabic language, and NVIDIA Riva played a significant role in improving the application’s performance. Using the NeMo toolkit in Riva, we were able to fine-tune an Arabic speech-to-text model to get a word error rate as low as 7.84 percent and reduced the training time of the model from days to hours using GPUs. We look forward to integrating these models in Riva’s end-to-end pipeline to ensure real-time latency.
— Karim Beguir, CEO and Co-founder at InstaDeep
Through the NVIDIA Riva early access program, we’ve been able to power our conversational AI products with state-of-the-art models using NVIDIA NeMo, significantly reducing the cost of getting started. Riva speech recognition has amazingly low latency and high accuracy. Having the flexibility to deploy on-prem and offer a range of data privacy and security options to our customers has helped us position our conversational AI-enabled products in new industry verticals.
— Rajesh Jha, CEO of Siminsights
At MeetKai, we build virtual assistants that make people's lives easier. When we started our company, we faced engineering and production challenges because there weren’t many high-quality, open-source conversational AI toolkits. NVIDIA NeMo helped our engineering efforts by providing easy-to-use APIs and reducing our costs by 25 percent. We look forward to continuing to work with NeMo to create the ultimate AI helper.
— James Kalpan, CEO of MeetKai
Kensho leverages S&P Global's world-class data and research to build amazing tools that help people make fact-based decisions. Using NVIDIA NeMo on GPUs, Kensho successfully transcribed tens of thousands of earnings calls, management presentations, and acquisition calls, unlocking double-digit accuracy improvements and enabling S&P Global to increase earnings call coverage by more than 25 percent.
— Keenan Freyberg, Product Manager at Kensho
Our goal with SpeechBrain at MILA is to build an all-in-one toolkit that can significantly speed up research and development for speech models. We’re interested in pushing the boundaries of speech technologies even further by integrating with NeMo modules, particularly speech recognition and language modeling.
— Mirco Ravanelli, Speech and Deep Learning Scientist at MILA
Easily Compose New Model Architectures
NeMo includes domain-specific collections for ASR, NLP and TTS to develop state-of-the-art models such as Citrinet, Jasper, BERT, Fastpitch, and HiFiGAN. A NeMo model is composed of neural modules, which are the building blocks of models. The inputs and outputs of these modules are strongly typed with neural types that can automatically perform the semantic checks between the modules.
NeMo is designed to offer high flexibility and you can use the Hydra framework to modify the behavior of models easily. For instance, you can modify the architecture of the Citrinet Encoder module in the following diagram using Hydra.
Train State-of-the-Art Conversational AI Models
Several NeMo pre-trained, state-of-the-art models in NGC are trained for over 100,000 hours on NVIDIA DGX™ across open-source, free datasets. You can fine-tune these models or modify them with NeMo before training for your use case.
NeMo uses mixed precision on Tensor Cores to speed up training upto 4.5X on a single GPU versus FP32 precision. You can further scale training to multi-GPU systems and multi-node clusters.
Large-Scale Language Modeling with NeMo Megatron
Large-scale, Transformer-based language models are being applied to a broad set of natural language tasks such as text generation, summarization, and chatbots.
NeMo Megatron is an end-to-end platform that delivers high training efficiency across thousands of GPUs and makes it practical for enterprises to deploy large-scale NLP. It provides capabilities to curate training data, train large-scale models up to trillions of parameters and deploy them in inference.
It performs data curation tasks such as formatting, filtering, deduplication, and blending that can otherwise take months. It includes state-of-the-art parallelization techniques such as tensor parallelism, pipeline parallelism, sequence parallelism, and selective activation recomputation, to scale models efficiently.
These models can be exported to the NVIDIA Triton™ Inference Server to run large-scale NLP models on multiple GPUs and multiple nodes.
Using ready-to-use recipes and the hyperparameter tool, which selects the best compute configuration for your model size, you can get started with building and deploying LLMs easily.
NeMo Megaton is optimized to run on DGX system on-prem as well as in the cloud.
Learn more about NeMo Megatron
Apply for Early Access
Flexible, Open-Source, Rapidly Expanding Ecosystem
NeMo is built on top of PyTorch and PyTorch Lightning, providing an easy path for researchers to develop and integrate with modules with which they are already comfortable. PyTorch and PyTorch lightning are open-source python libraries that provide modules to compose models.
To provide the researcher's flexibility to customize the models/modules easily, NeMo integrated with the Hydra framework. Hydra is a popular framework that simplifies the development of complex conversational AI models.
NeMo is available as an open-source so that researchers can contribute to and build on it.
Deploy to Production using NVIDIA AI
To leverage the NVIDIA AI platform and deploy NeMo speech models in production with NVIDIA Riva, developers should export NeMo models to a format compatible with Riva and then execute Riva build and deploy commands for creating an optimized skill that can run in real-time.
The documentation includes detailed instructions for exporting and deploying NeMo models to Riva.
NeMo is built on top of the popular PyTorch framework and facilitates researchers to use the NeMo modules with PyTorch applications.
NeMo with Pytorch Lightning enables easy and performant multi-GPU/multi-node mixed-precision training.
Data Generation and Data Annotation Partners
NVIDIA NeMo provides the capability to train and fine tune state-of-the-art models built using it. Fine-tuning models requires high quality labeled data, which might not be readily available. NeMo is integrated with several easy-to-use speech and language data labeling tools to help acquire labeled data as well as label custom data.
Get Started with Tutorials
Check out tutorials to get up and running quickly with state-of-the-art speech and language models.
Take a NeMo Tour
Understand the advantages of using NVIDIA NeMo with a Jupyter Notebook walkthrough.