...
- Goal is to support ML models in non-real time and near-real time RIC usecases.
- quickly import an Acumos model into RIC and adapt it into as an xApp (near-real time).
- deploy Acumos models as is into non-real time (mostly on ONAP side).
- Priority is to get something working with minimal changes possible on ML models
- focus on performance in the later releases, since many ML models take some time to execute anyway.
- Build a standard xApp/Acumos microservice adapter
- deployed along with the Acumos ML model in one Kubernetes pod.
- Adapter speaks RMR protocol to RIC
- communicates with the Acumos ML model in the standard http / GRPC manner.
- Configuration needed for each deployment
- to tell adapter how to speak with Acumos ML model.
- can be auto generated using ML model protobuf definition.
- Consider writing custom RMR model runner
- for performance in near-real time RIC xApps in the following releases.
Gerrit Repo: https://gerrit.o-ran-sc.org/r/admin/repos/ric-app/ml
Slide deck with Acumos intro and ML based xApp design for Amber release
| View file | ||||
|---|---|---|---|---|
|
Gerrit Repo
https://gerrit.o-ran-sc.org/r/admin/repos/ric-app/ml
Why Acumos?
In short, the unique features are 1) Distributed AI Marketplace, 2) Interoperable ML Microservices, and 3) Common Open Source Framework to AI.
...