OCUDU AI-RAN Release 4 : Hierarchical AI-RAN for Dynamic Network Slicing
Slice-ML implements intelligent RAN resource slicing using hierarchical deep learning within the OCUDU-RAN Distributed Unit, enabling dynamic allocation for services with varying performance needs. The system operates at two timescales: a slower LSTM-based controller adjusts physical resource block (PRB) allocation between enhanced mobile broadband (eMBB) and ultra-reliable low latency communication (URLLC) slices, and a faster attention-based model prioritizes user equipment (UE) scheduling within each slice. Separate models are used for eMBB and URLLC, while the existing DU resource allocator handles final assignment. This lightweight C++ implementation performs ML inference natively without introducing framework overhead into the real-time scheduling path, supporting data collection, shadow operation, and runtime safety features with conventional scheduling as a fallback. Validated models can be hot-swapped without DU restarts, allowing for controlled deployment and continuous improvement.