AMD AI Workbench allows users to deploy custom models not found in the AIM Catalog, such as those from Hugging Face Hub or internally trained models. The tool integrates with the AMD Inference Microservice (AIM) framework, enabling seamless deployment and management of models as inference endpoints. Users can import models via the GUI, configure runtime settings, and deploy them as part of their projects. This capability expands the range of AI models available for deployment on AMD hardware, supporting both public and gated models with appropriate access credentials. Source: amd
The AI Workbench streamlines the process of onboarding custom models by guiding users through three key steps: specifying the model source, confirming model information, and setting the runtime profile. During the model import, users can preview repository details, adjust display properties, and ensure the model is compatible with the AIM Engine. If the model is gated, a Hugging Face token is required to access its metadata and weights. The tool also automatically configures the runtime environment, detects accelerators, and selects performance profiles, abstracting the complexity of model serving. Source: amd
The blog outlines the workflow for deploying a model from Hugging Face, including adding project secrets for storage access and navigating the Models page to import and deploy the model. Custom models are scoped to the project they are imported into and can be deployed as inference endpoints with OpenAI-compatible APIs. The process includes verifying storage credentials, importing the model, and monitoring its deployment status. The blog emphasizes the importance of checking GPU quotas and ensuring model compatibility with the allocated resources. Source: amd
