Energy Efficient Scheduling of AI/ML Workloads on Multi Instance GPUs with Dynamic Repartitioning

This is an arXiv CS digest (multiple new submissions) covering systems and ML infrastructure research. Key highlights: a dynamic-repartitioning scheduler for NVIDIA Multi-Instance GPUs (MIGs) that uses simulations and reinforcement learning to cut energy and tardiness vs. static/no partitioning; an analytical and empirical study of distributed speculative decoding (edge draft model + cloud target) that shows limited per-request latency benefit over co-located speculative decoding except in low-RTT/multi-tenant scenarios; an architecture and real-cluster experiments for grid-responsive, power‑/