Efficient and Portable Support for Overdecomposition on Distributed Memory GPGPU Platforms

Collection of recent arXiv papers (May 2026) across distributed systems, ML systems, and HPC. Highlights include: techniques to support overdecomposition efficiently and portably on heterogeneous GPGPU platforms; new lower/upper-bound results showing the LOCAL-model complexity on trees depends on nodes' knowledge of network size; communication-efficient Byzantine agreement algorithms that reduce prediction-exchange overhead (unauthenticated and authenticated variants); heterogeneous CPU+GPU solvers implemented in SYCL; PipeSD, a cloud-edge speculative-decoding pipeline for faster LLM inference