Recently, ByteDance announced the launch of Klever, an open source machine learning platform . Klever is a cloud native machine learning platform that supports OCI (Open Container Initiative) standard storage training models and supports online model service deployment. Algorithm scientists can use Klever for model management, model analysis, model conversion, and model services. It has solved the following problems in the intelligent technology landing process:
- Model management and distribution
- Model analysis and conversion
- Online model service deployment and management
Klever has four self-developed components:
- ormb: model packaging, decompression, upload, download tool ( click to learn about ormb )
- model-registry: model warehouse and model service API management layer
- modeljob-operator: ModelJob controller, which manages model analysis and model conversion tasks
- klever-web: front-end components
And rely on three open source components:
- Istio: open source service grid components, model services expose model service addresses to the outside through Istio, so that model services can be distributed by content and proportionally.
- Harbor: Model bottom storage component, hierarchical storage of model configuration and model files
- Seldon Core: The controller of Seldon Deployment CRD for open source model service management, and realizes model service management through SeldonDeployment CR
