AMD AI Workbench#
This page introduces the AMD AI Workbench, summarizes its main capabilities, and points you to installation, concepts, and hands-on guides.
Solution overview#
The AMD AI Workbench is an interface for developers to easily manage the lifecycle of their AI stack. The Workbench provides an easy-to-use low-code option for running and managing AI workloads. Key capabilities include one-click deployment of AIMs from the AIM Catalog, plus bundled tools such as Jupyter Notebooks, Visual Studio Code and MLOps integrations including MLflow for experiment tracking and model lifecycle management.
The AMD AI Workbench can be deployed standalone, or combined with AMD Resource Manager. The core AI development features (model deployment, fine-tuning, workspaces, chat) are available in all deployment modes. When integrated with AMD Resource Manager, additional enterprise features such as quota management, organizational hierarchy, and advanced access control become available. See the Installation guide for deployment options and installation details.
Where to Start#
Installation
Deployment modes and installation guides.
Concepts
Core areas of the Workbench UI and how they fit together.
Quick start
Guided tutorials for deploying models, fine-tuning, and workspace-based development.
Key Features#
The AMD AI Workbench includes the following capabilities:
AIM Catalog#
The AMD AI Workbench offers a comprehensive catalog of AMD Inference Microservices (AIMs), which are optimized inference containers that provide standardized, production-ready services for state-of-the-art large language models (LLMs) and other AI models. Each AIM includes built-in model caching and hardware optimization for AMD accelerators (AMD Instinct™ GPUs and AMD EPYC™ CPUs). Developers can easily discover, deploy, and fine-tune compatible models for their AI use cases.
AI Workspaces#
The AMD AI Workbench provides developers with tools and frameworks and easy access to GPU resources to accelerate AI development and experimentation, featuring a comprehensive catalog of optimized AI workloads and models for AMD compute. The workloads include the most common developer tools, such as Jupyter Notebooks, Visual Studio Code, PyTorch, and TensorFlow.
Training and fine-tuning#
Fine-tuning a model allows developers to customize it for their specific use case and data. AMD AI Workbench provides a certified list of base models that developers can fine-tune, and allows customization of certain hyperparameters to achieve the best results.
Chat and compare#
The chat page allows developers to experiment with models they have access to. Developers can modify generation parameters to see how they affect the model’s response. The model comparison view allows developers to compare the output of different models using the same settings.
GPU-as-a-Service#
The AMD AI Workbench provides developers with self-service access to workspaces with GPU resources. Platform admins can set project quotas for GPU usage so teams always have the right amount of resources available.
API Keys for Programmatic Access#
API keys provide secure programmatic access to deployed AI models. Developers can create API keys to integrate deployed models into applications, automate inference workflows, or access endpoints from external systems. API keys support configurable expiration times, renewal, and fine-grained access control by binding them to specific model deployments.
For direct access to the AI Workbench management API (projects, workloads, secrets, datasets, and related resources), developers authenticate with Keycloak / OAuth2 bearer tokens.
Secrets#
Manage credentials such as Hugging Face tokens through Secrets and project configuration.
Running AI workloads through the command-line#
Developers can also deploy and run AI workloads through the command-line interface using kubectl. Visit AI workloads for workloads that are pre-validated, open-source and continuously updated.
Terminology reference#
Term |
Explanation |
|---|---|
AIM |
AMD Inference Microservice, a containerized, profile-driven inference service for a model. |
AIM catalog |
The set of published AIM images you can pull, deploy, and fine-tune from the Workbench. |
Workspace |
A GPU-backed development environment (for example JupyterLab or VS Code) tied to your project. |
Project |
In combined mode, an isolation boundary with namespace, quota, and membership (managed by AMD Resource Manager). In standalone mode, a single implicit project namespace is used. |
Quota |
Guaranteed share of cluster resources (for example GPUs) for a project when Resource Manager is enabled. |
API key |
Token used to call deployed model HTTP APIs from applications. |
Secret |
Stored credential (for example Hugging Face token) used by workloads and catalog actions. |