Release Notes#
Version: 1.5.2#
Release Date: November 16, 2025
Documentation#
For details, see documentation on this site (https://enterprise-ai.docs.amd.com).
New Features and Updates: General#
Public Open Source Repositories
Installation with Helm charts for all components
New Features and Updates: AMD Inference Microservices (AIMs)#
Catalog of inference containers of popular open weights models
Run with Docker or deploy on Kubernetes platforms. Instructions for scalable deployment with KServe.Optimized for AMD Instinct MI300X
Optimized inference containers for most popular models, supplemented with a larger selection of models validated to perform well.
New Features and Updates: Solution Blueprints#
Catalog of example applications using AIMs: to facilitate getting started with development.
New Features and Updates: AI Workbench#
Low Code UI for managing AIM deployments: Deploy and interact with AIMs from the catalog.
API key management: Share API keys with end users for controlling access while allowing integration of AIM inference in 3rd party applications.
Monitoring Dashboard: Monitor utilization and performance of deployed AIMs.
New Features and Updates: AMD Resource Manager#
Bucket storage management: Ability to connect external bucket storage to a project.
Known Issues#
First request to each AIM fails due to timeout. Subsequent requests generates responses as expected.
ComfyUI takes 10-15min to launch first time
Model download may occasionally fail due to connectivity problems with the HuggingFace model registry or issues with the MinIO bucket storage connection.
Cluster storage is not auto-monitored, which may result in storage becoming full and causing workload failures.
Certain models, like the Llama 3.2 Vision models, may not be available for download and deployment due to geographical restrictions. This can cause the workloads to fail.
Only one cluster is supported: the GUI suggests that another cluster can be onboarded, but this is not yet fully supported.
ConfigMaps/Secrets/HttpRoutes in the YAML file are not treated as part of the workload and will not be automatically deleted if the workload is deleted.
Custom Resources have only partial support; deleting the workload from the Resource Manager will not delete the custom resource.
Workloads submitted via kubectl will not be displayed for team members on the AI workbench; they are only displayed for platform admins on the Resource Manager.
Version: 0.5.2#
Release Date: October 2, 2025
This is an early access release of the AMD Resource Manager and AMD AI Workbench.
Documentation#
For details, see documentation https://enterprise-ai.docs.amd.com.
New Features and Updates: AMD Resource Manager#
Kubernetes Cluster Onboarding: Installation and configuration of Enterprise ready Kubernetes cluster for AMD GPUs
Compute Quota Management: Create projects with assigned users and allocated compute resource (GPU) quota
Workload Management: Efficient management of both Deployments (inference & workspaces) and Jobs (fine-tuning)
Workload Management: Use AI Workbench or kubectl to submit workloads to your project namespace
Monitoring: Dashboards for monitoring compute utilization on cluster, project and workload levels
New Features and Updates: AI Workbench#
Models and Data Catalogues: Catalogue with popular open source models and ability to upload fine-tuning data
Low Code UI for Inference and Fine-tuning: Fine-tuning recipes for selected models (e.g. Llama 3.1) and Chat UI for model interaction
Inference API: Deploy models for inference behind API Gateway, for both community models and custom models fine-tuned on the platform
Workspaces: Deploy VSCode and JupyterLab workspaces with selected image and pre-allocated compute resources
Workspaces: Deploy ComfyUI for image generation or Mlflow for experiment tracking
Reference Workloads for kubectl users: Get started with customizing your own workloads from our open source catalogue of reference workloads (see docs for details)
Known Issues#
Model download may occasionally fail due to connectivity problems with the HuggingFace model registry or issues with the MinIO bucket storage connection.
Cluster storage is not auto-monitored which may result in storage becoming full and causing workload failures.
Certain models, like the Llama 3.2 Vision models may not be available for download and deployment due to geographical restrictions. This can cause the workloads to fail.
When downloading models from the Community Models page, the state of the UI card might not update to “Deploy” A page refresh is then needed.
Only one cluster is supported the GUI suggests that another cluster can be onboarded, but this is not yet fully supported.
ConfigMaps/Secrets/HttpRoutes in the YAML file are not treated as part of the workload and will not be automatically deleted if the workload is deleted
Custom Resources have only partial support, deleting the workload from the Resource Manager will not delete the custom resource
Workloads submitted via kubectl will not be displayed on the AI workbench they are only displayed on the Resource Manager
Release Date: September 15, 2025#
Documentation#
For details, see documentation https://enterprise-ai.docs.amd.com.
New Features: AMD Resource Manager#
Kubernetes Cluster Onboarding: Installation and configuration of Enterprise ready Kubernetes cluster for AMD GPUs
Compute Quota Management: Create projects with assigned users and allocated compute resource (GPU) quota
Workload Management: Efficient management of both Deployments (inference & workspaces) and Jobs (fine-tuning)
Monitoring: Dashboards for monitoring compute utilization on cluster, project and workload levels
New Features: AI Workbench#
Models and Data Catalogues: Catalogue with popular open source models and ability to upload fine-tuning data
Low Code UI for Inference and Fine-tuning: Fine-tuning recipes for selected models (for example Llama 3.1) and Chat UI for model interaction
Inference API: Deploy models for inference behind API Gateway, for both community models and custom models fine-tuned on the platform
Workspaces: Deploy VSCode, JupyterLab or ComfyUI workspaces with pre-allocated compute resources
Known Issues#
Delays in JupyterLab Workspace It takes about 30 seconds after the launch button gets activated before the JupyterLab can actually be launched.
Community Models document link is not pointing to right document after download
Custom Models “Failed” models have “deploy” action menu item. They shouldn’t, since you cannot deploy them.
Model download may occasionally fail due to connectivity problems with the HuggingFace model registry or issues with the MinIO bucket storage connection.
Cluster storage is not auto-monitored which may result in storage becoming full and causing workload failures.
Certain models, like the Llama 3.2 Vision models may not be available for download and deployment due to geographical restrictions. This can cause the workloads to fail.
When downloading models from the Community Models page, the state of the UI card might not update to “Deploy” A page refresh is then needed.
Only one cluster is supported the GUI suggests that another cluster can be onboarded, but this is not yet fully supported.
Kubectl usage for workload deployment needs workaround the AI Resource manager currently does not manage those workloads. This is coming in next release. If you need this now, we can share a workaround that may be sufficient for some teams of cooperative users
The usage of SMTP based functionality to invite users requires the following additional configuration, in addition to the configuration highlighted here:
In the API deployment, the POST_REGISTRATION_REDIRECT_URL needs to be changed to
https://airmui. instead ofhttps://airm.In the keycloak AIRM Admin API Client, the redirectUris needs to include the new POST_REGISTRATION_REDIRECT_URL with a * at the end.