HypeReca: Hybrid Partitioned Embeddings for Recommendor Model Training Acceleration

This repository includes the core code of ATC'25 paper SpaRes: Distributed Heterogeneous In-Memory Embedding Database for Training Recommender Models.

The project was originally called DrESS: Distributed Embedding Serving System. We currently keep the name dress in our code.

Usage

Use cmake to build this project into libdress.so. A variant of HugeCTR is an example of using HypeReca.