heterogeneous computing

New Reinforcement Learning Framework WeCAN Improves DAG Scheduling Efficiency in Heterogeneous Environments

Researchers propose WeCAN, an end-to-end reinforcement learning framework for scheduling directed acyclic graphs (DAGs) on heterogeneous resources. The method uses a two-stage single-pass design with a weighted cross-attention encoder and a skip-extended generation map to close optimality gaps. Experiments on TPC-H query DAGs and ML-compiler graphs show improved makespan with inference time comparable to classical heuristics.