Red Hat has updated your machine learning and Artificial Intelligence platform OpenShift AIcon improvements focused on AI flexibility and scalability. Red Hat OpenShift AI 2.15 will feature optimized tuning and tracking capabilities for building and deploying AI-ready applications at scale in hybrid cloud environments, among other new features.
With all of them, platform users will have an easier time accelerating their AI and machine learning work, as well as achieving more consistency in their operations. Also with more regularity and more robust security in public clouds, data centers and edge environments.
Among the advanced functions offered by Red Hat OpenShift AI 2.15 is model registration, although for now in the technological testing phase. It is a centralized repository of models, in which those that are registered can be viewed and managed.
It allows you to share, version, deploy and monitor generative and predictive AI models, as well as metadata and model artifacts, in an organized and structured way. Additionally, it gives the opportunity to configure several model registers.
Another of its novelties is the detection of data deviation, which aims to monitor changes in the distribution of input data for the machine learning models deployed. With it, data scientists can detect when the real-time data used for model inference significantly deviates from the training data. This results in more reliable models, thanks to the fact that they will be in line with the real data and it will be possible to maintain the precision of their predictions over time.
OpenShift AI will also now include improvements to bias detection tools, dedicated to checking the fairness and fairness of models, in order to improve confidence in them. These tools, in addition to reporting potential biases in training data, also monitor the fairness of models when deployed in real environments.
Bias detection tools are integrated into the platform from the open source TrustyAI community, which will offer a package to develop and deploy AI responsibly.
Efficient tuning with low range adapters (LoRA) allows large language models to be tuned more efficiently. Thus, companies have an easier time scaling AI workloads, while reducing costs and resource consumption.
On the other hand, performance and flexibility are also improved thanks to the optimization of training and adjustment of models in cloud-native environments. And support for the Nvidia NIM interface microservices suite, for its ease of use, leads to accelerated delivery of generative AI applications.
The integration with NIM comes out of the Nvidia AI Enterprise platform, and will help accelerate generative AI deployments. It achieves this in part by supporting a wide range of AI models to deliver scalable inference, both on-premises and in the cloud, through APIs.
Now supporting AMD GPUs, Red Hat OpenShift AI provides access to an AMD ROCm workbench image for developing models with AMD GPUs. It also provides access to images for service and training use cases with these GPUs.
Red Hat OpenShift AI 2.15 will also feature new features for serving generative AI models, including the vLLM service runtime for KServe. This capability incorporates the open source model service runtime for deploying large language models. In addition, they allow users to add personality options to the platform depending on their needs.
The platform will also now be compatible with KServe Model cars, and will incorporate Open Container Initiative (OCI) repositories as an option to store and access versions of models in containers. As for choosing private or public routes for endpoints in KServe, it makes it easy to harden the security of the model, and direct them to internal endpoints when needed.
OpenShift AI 2.15, which will be available from the middle of this monthimproves data science pipelines and experiment tracking, in addition to facilitating the management, comparison and analysis of pipeline executions grouped in a logical structure.
It also integrates hyperparameter tuning with Ray Tune, including advanced optimization algorithms for improving the accuracy and efficiency of training predictive and generative AI models.
