Hugging Face and Amazon SageMaker Studio Integration

Hugging Face and Amazon SageMaker Studio Integration

Hugging Face and Amazon SageMaker AI have launched a deep-link integration that enables developers to transition from model discovery on Hugging Face to hands-on experimentation in SageMaker Studio via a single selection. This integration removes the manual friction of configuring domains, IAM permissions, and GPU quotas, creating a direct path from open-model discovery to enterprise deployment.

Streamlined Workflow via Deep Links

Developers can now use action buttons on supported Hugging Face model pages to launch specific SageMaker Studio workflows without needing to search for the model again once inside the AWS environment. These deep links preserve the model context across platforms:

  • Customize on SageMaker AI: Directly opens the Model Customization page in Studio with the selected model pre-loaded for fine-tuning.
  • Deploy on SageMaker AI: Directly opens the Deployment page in Studio with the model pre-configured for endpoint deployment.

Automated Environment and Permission Setup

To eliminate manual configuration, new SageMaker Studio environments created through this integration automatically provision a domain with pre-configured permissions.

Managed Permissions Policy

A new managed policy, AmazonSageMakerModelCustomizationCoreAccess, is automatically attached to the environment. This policy provides the necessary permissions for serverless model customization jobs, including:

  • Supervised Fine-Tuning (SFT)
  • Direct Preference Optimization (DPO)
  • Reinforcement Learning with Verifiable Rewards (RLVR)
  • Reinforcement Learning from AI Feedback (RLAIF)

These jobs can be deployed to either SageMaker AI or Amazon Bedrock endpoints. For users with existing Studio environments, the system provides actionable messages and links to documentation to help add these permissions manually.

Integrated GPU Quota Visibility

SageMaker Studio now surfaces GPU quota availability directly within the instance selection list, removing the need to navigate to the separate Service Quotas page. Users can immediately see the availability of G5 and G6 GPU instance types under their account limits. If a limit increase is required, the UI provides a direct redirect to the Service Quotas page for the specific instance type.

Implementation Walkthrough

The process for moving from Hugging Face to a working SageMaker endpoint involves four primary steps:

  1. Discover and Select: On a supported Hugging Face model page, the user selects either "Deploy on SageMaker AI" or "Customize on SageMaker AI".
  2. Sign-in: The user authenticates via the AWS Management Console. This step is skipped if an active session already exists.
  3. Land in Studio: The user arrives at the pre-configured Model Customization or Deployment page in SageMaker Studio with the model pre-selected.
  4. Test: Once the endpoint is deployed or the model is customized, inference can be tested directly using the Studio endpoint testing interface.

"Going from an open model on Hugging Face straight into SageMaker Studio in a single click, then fine-tuning or deploying it inside your own AWS environment with nothing to wire up, is the kind of experience open models have been missing," says Mark McQuade, Founder and CEO of Arcee AI.

Sources