EXaCTz: Guaranteed Extremum Graph and Contour Tree Preservation for Distributed- and GPU-Parallel Lossy Compression

This feed aggregates systems and ML-systems research with multiple operational security implications. Highlights include EXaCTz — a high-throughput GPU/distributed algorithm that enforces topological consistency in lossy-compressed scientific scalar fields; ModTrans — a translator to import real-world models into a distributed-training simulator; DWDP — a distributed weight-data-parallel LLM inference scheme that fetches remote experts on demand; several cloud/HPC optimization and GPU power/utilization prediction works that rely on telemetry (Slurm, NVIDIA DCGM); an anonymity-deanonymization理论