February 07, 2025

Release notes and fixed issues for version 2.0.47-b345.

New Features / Improvements🔗

Cloudera AI Workbench

Support is now provided for API keys to invoke applications deployed using Cloudera AI Workbenches. This not only eases the invocation of those applications programmatically but also allows one application to easily invoke another application that they have access to.
MLFLOW upgrade for Cloudera AI Workbenches now enables making use of the latest offerings and APIs from the MLFLOW community like evaluateLLM.

Cloudera AI Platform

The autoscaling range of Suspend Workflow is now set to the value 0 to ensure that other Kubernetes deployments outside the scope of MLX can deploy their pods on worker nodes.

Cloudera AI Registry

An enhanced error message is now displayed during model upload failure.
UI for Registered Models displays the environment name of the registry along with an error message when any user is unable to access any Cloudera AI Registry.
A checkbox is now added to enable Public Load Balancer for new Cloudera AI Registries on Azure.

Cloudera AI Inference service

The Hugging Face model server backend has been upgraded, which expands the compatibility with a larger number of model families, such as Llama 3.3 and models derived from it.
Llama 3.2 Vision Language Model NIM version has been updated to address compatibility with A10G (g5.*) and L40S (g6e.*) GPU instances on AWS.
You can now upgrade Cloudera AI Inference service using the UI. Previously, the upgrade was supported only using CDPCLI.
You can now upgrade from Cloudera AI Inference service version 1.2.0-b80 to version 1.3.0-b113 or higher. Note that you cannot upgrade from 1.3.0-b111 to 1.3.0-b113 or higher. For more information on the 1.3.0-b111 upgrade issue and workaround, see the Known Issues section.

Cloudera AI Workbench

Previously, due to an issue, users could stop sessions under projects that they were not authorized to access using the session’s UUID. This issue is now resolved. (DSE-39798)
Previously, when a Kubernetes object was deleted, and the reconciler was overwhelmed by a large number of events, the Deleted status failed to propagate properly. This issue is now resolved. (DSE-41431)
Previously, the stopped_at column was not correctly populated when applications were stopped. This issue is now resolved. (DSE-41636)
Previously, engine pods were stuck in the Init:StartError state and you had to manually delete it. With this fix, pods stuck in Init:StartError in the Garbage Collection will be deleted after a certain grace period. (DSE-41430)
Previously, Spark environment configurations were not inherited by models running Spark. With this fix, models use the appropriate Spark configurations to run Spark. (DSE-36343)

Cloudera AI Registry

An issue around how Hugging Face token was being consumed during the import of a model was addressed. (DSE-41714)
The Cloudera AI Registry deletion flow is improved to take care of race conditions when both creation and deletion are triggered in a short frame of time. (DSE-41634)

Cloudera AI Inference service

Previously, the GetEndpointLogs failed with an error. With this fix, endpoint logs for the model container do not exceed the gRPC messaging size. (DSE-41765)
A new field called loadBalancerIPWhitelists is added to display a list of IPs whitelisted for the load balancer and deprecated isPublic and ipAllowlist. (DSE-39397)
Infrastructure nodes are no longer shown as instances that can be used for deploying a new endpoint. (DSE-41726)

ML Runtimes

Previously, due to an issue, to ensure the compatibility of AMPs with ML Runtimes 2025.01.1, users had to switch to JupyterLab PBJ Workbench in the AMPs’ .project-metadata.yaml file or use jobs instead of sessions for automated tasks. This issue is now resolved. (DSE-41263)
Resolved issues related to using R interactively in PBJ Runtimes. (DSE-41771)