NVIDIA and Hugging Face Offers Inference-as-a-Service with 5x Token Efficiency for AI Models

Open Source platform Hugging Face is offering developers Inference-as-a-Service that will be powered by NVIDIA’s NIM. The new service provides 5x better token efficiency for AI models and allows immediate access to NIM microservices running on NVIDIA DGX Cloud. 

The new inference-as-a-service was announced at the ongoing SIGGRAPH 2024, a premier conference and exhibition on computer graphics and interactive techniques happening at Denver, Colorado. The new service will facilitate developers to deploy powerful LLMs such as Llama 2, Mistral AI models and many more with optimisation from NVIDIA NIM microservices. Hugging Face Enterprise Hub users can access serverless inference for increased flexibility and minimal infrastructure overhead with NVIDIA NIM.

When accessed as a NIM, large models such as the 70-billion-parameter version of Llama 3, will deliver up to 5x higher throughput when compared with off-the-shelf deployment on NVIDIA H100 Tensor Core GPU-powered systems.

The new inference service supports Train on DGX Cloud, an AI training service that is already available on Hugging Face.

The Omnipresent NVIDIA NIM 

NVIDIA NIM is a set of AI microservices, including NVIDIA AI foundation models and open-source community models, that has been optimised for inference with standard APIs. It improves token processing efficiency and enhances the NVIDIA DGX Cloud infrastructure, accelerating AI applications. This setup provides faster, more robust results. 

The NVIDIA DGX Cloud platform is tailored for generative AI, offering developers reliable, accelerated computing infrastructure for faster production readiness. It supports AI development from prototyping to production without requiring long-term commitments. 

Hugging Face Dominates

The new announcement banks on an existing partnership between both tech companies and is only going to foster the developer community further. Interestingly, Hugging Face recently announced its profitability with a 220-member team. They also released SmolLM, a series of small language models.