Chapter 2. New features and enhancements
This section describes new features and enhancements in Red Hat OpenShift AI 2.22.
2.1. New features Copy linkLink copied to clipboard!
- Model-serving runtimes display version information
With this update, you can now view the version of each model-serving runtime. The runtime version is displayed in the following locations in the dashboard:
- On the Serving runtimes page.
- In Settings, under Serving runtimes.
- On the Models and model servers page, in the Serving runtimes column.
Custom runtimes do not display a version automatically. To display the runtime version for a custom runtime, add opendatahub.io/runtime-version
to the ServingRuntime
object as an annotation.
- Ability to select access mode for storage classes
- When adding cluster storage to a data science project or workbench, you can now select the access mode for the storage class based on administrator configurations. Access mode is a Kubernetes setting that defines how nodes can access a volume. This update improves flexibility and efficiency in managing shared data for workbenches. Previously, all storage classes defaulted to ReadWriteOnce (RWO), which prevented multiple users from sharing the same storage class for collaborative work.
2.2. Enhancements Copy linkLink copied to clipboard!
- Updated vLLM component versions
OpenShift AI supports the following vLLM versions for each listed component:
- vLLM CUDA v0.9.0.1 (designed for FIPS)
- vLLM ROCm v0.8.4.3 (designed for FIPS)
- vLLM Power v0.9.1
- vLLM Z v0.9.1 (designed for FIPS)
vLLM Gaudi v0.7.2.post1
For more information, see
vllm
in GitHub.
- Support added for LLM-as-a-Judge metrics
You can now use LLM-as-a-Judge metrics with LM-Eval in TrustyAI. Large language models (LLMs) can be used as human-like evaluators to assess the quality of outputs from another LLM that are not easily quantifiable, such as rating a piece of creative writing. This is known as LLM-as-a-Judge (LLMaaJ).
See LM-Eval scenarios for example evaluations.
- Distributed workloads: additional training images tested and verified
The following additional training images are tested and verified:
CUDA-compatible Ray cluster image
A new Ray-based training image,
quay.io/modh/ray:2.46.0-py311-cu121
is tested and verified. This image is compatible with AMD accelerators that are supported by CUDA 12.1.ROCm-compatible Ray cluster image
The ROCm-compatible Ray cluster image
quay.io/modh/ray:2.46.0-py311-rocm62
is tested and verified. This image is compatible with AMD accelerators that are supported by ROCm 6.2.
These images are AMD64 images, which might not work on other architectures. For more information about the latest available training images in Red Hat OpenShift AI, see Red Hat OpenShift AI Supported Configurations.
- Improved reliability for the OpenShift AI Operator
- The OpenShift AI Operator now runs with three replicas instead of a single instance, which improves resilience and reliability for production workloads. This enhancement reduces interruptions on OpenShift AI services and distributes webhook operations across multiple instances.
- Support for Kubeflow Pipelines 2.5.0 in data science pipelines
- Data science pipelines have been upgraded to Kubeflow Pipelines (KFP) version 2.5.0. For more information, see the Kubeflow Pipelines release documentation.
- Automated creation of Elyra resources by the notebook controller
Previously, the
elyra-pipelines-<notebook-name> RoleBinding
andds-pipeline-config Secret
resources were provisioned by the dashboard component, which lacked integration with the controller’s lifecycle management. This dependency also required you to deploy the OpenShift AI dashboard, even when you only needed the pipeline functionality.With this release, the notebook controller now automatically creates these resources, allowing you to use workbenches and pipelines independently of the dashboard component. This change simplifies setup and ensures more consistent lifecycle management.
- Seldon MLServer version 1.6.1 runtime now tested and verified
Red Hat has tested and verified the Seldon MLServer version 1.6.1 runtime, improving compatibility with popular predictive AI models. The following models were tested for KServe (REST and gRPC):
- Scikit-learn
- XGBoost
- LightGBM
- CatBoost
- MLflow
- Hugging Face
- Operator dependencies for model registry removed
The Red Hat Authorino, Red Hat OpenShift Serverless, and Red Hat OpenShift Service Mesh Operators are no longer required to use the model registry component in OpenShift AI.
Existing model registry instances will automatically be migrated to use OpenShift OAuth proxy authentication. New model registry instances created from the OpenShift AI dashboard will use OAuth proxy by default. New instances created with the older
v1alpha1
API and Istio configuration will automatically be updated to use OAuth proxy.Existing authorization configuration for older model registry instances such as Kubernetes RBAC resources will continue to work as expected.