Hugging Face and Microsoft Strategic Collaboration Update May 2024

Hugging Face and Microsoft have deepened their strategic collaboration to simplify the deployment and use of open-source AI models across cloud infrastructure, hardware, and developer tools. This partnership focuses on making open models more accessible through the Azure Model Catalog, optimizing performance on new AMD hardware, and streamlining the developer workflow between Hugging Face Spaces and VS Code.

Expanded Open Model Access in Azure AI Studio

Microsoft Azure AI Studio now features an expanded Hugging Face Collection, allowing for one-click deployment of several high-performance open Large Language Models (LLMs). This integration enables Azure AI customers to deploy models directly from the Hugging Face Hub or within the Azure AI Studio interface.

Newly added models to the collection include:

  • Llama 3 (Meta)
  • Mistral 7B (Mistral AI)
  • Command R Plus (Cohere for AI)
  • Qwen 1.5 110B (Qwen)
  • Various top-performing fine-tuned models from the Hugging Face community as ranked on the Open LLM Leaderboard.

Hardware Optimization with AMD MI300X on Azure

Azure has announced the general availability of Azure ND MI300X virtual machines (VMs) powered by AMD Instinct MI300 GPUs. Hugging Face has collaborated with AMD and Microsoft to optimize the performance and cost-efficiency of Hugging Face models on these VMs.

This optimization is achieved through the Optimum-AMD open-source library, which provides ROCm integrations and continuous testing of models and libraries on AMD Instinct GPUs to ensure high performance.

Advancements in Open Science and Local Inference

Microsoft continues to contribute to open science by hosting nearly 300 models on the Hugging Face Hub, including the MIT-licensed Phi-3 family. The Phi-3 mini model, featuring 3.8 billion parameters, is designed for on-device applications and is currently available in Hugging Face Chat.

To facilitate local inference, Hugging Face and Microsoft are leveraging WebGPU through Transformers.js and ONNX Runtime Web. This collaboration allows models like Phi-3 to run locally in the browser; demonstrations have shown Phi-3 generating over 70 tokens per second using WebGPU.

Enhanced Developer Experience with Spaces Dev Mode

A new integration called Spaces Dev Mode allows developers to connect their Hugging Face Spaces—where over 500,000 AI demo applications exist—directly to local VS Code environments or use a web-hosted VS Code instance.

Spaces Dev Mode is currently in beta and available to Hugging Face PRO subscribers, aiming to create a more seamless transition between local development and cloud-hosted AI demos.

Sources