Objective: Achieve orchestration of data operations and Kubernetes job scheduling systemsâ
Support temporality through Kueue
Once data migration is completed, run data preheating, triggering the running of machine learning tasks (such as tfjob, mpiJob, pytorchJob, sparkJob)
After computation is completed, data migration and cache cleaning can be carried out
Choose data access methods based on the scheduling results of the Kubernetes scheduler (default scheduler, Volcano, YuniKorn)
If scheduled to ordinary nodes with shared operating system kernels, adaptively use csi plugin mode
If scheduled to Kata container nodes with independent operating system kernels, you can use the sidecar mode adaptively and support scalable modifications by cloud vendors
Objective: Simplify the work of operation and maintenance and AI developers through Python SDKâ
Support basic data operation
Combine with Hugging face and Pytorch to support transparent data acceleration through pre-reading and multi-stream reading
Support defining automated data flow operations
Objective: Further deeply integrate the machine learning ecosystem to simplify the user experienceâ
Integrate with Kubeflow Pipelines to accelerate datasets in the pipeline
Integrate with Fairing for model development and deployment in the notebook environment
Integrate with KServe to facilitate model deployment