KubeDL enables deep learning workloads to run on Kubernetes more easily and efficiently.
KubeDL is a CNCF sandbox project.
- Support training and inferences workloads (Tensorflow, Pytorch. Mars etc.)in a single unified controller. Features include advanced scheduling, acceleration using cache, metadata persistentcy, file sync, enable service discovery for training in host network etc.
- Automatically tunes the best container-level configurations before an ML model is deployed as inference services. – Morphling Github
- Model lineage and versioning to track the history of a model natively in CRD: when the model is trained using which data and which image, each version of the model, which version is running etc.
- Enables storing and versioning a model leveraging container images. Each model version is stored as its own image and can later be served with Serving framework.
Check the website: https://kubedl.io
KubeDL-Morphling paper accepted at ACM Socc 2021: Morphling: Fast, Near-Optimal Auto-Configuration for Cloud-Native Model Serving