Efficient Training on Multiple Consumer GPUs with RoundPipe

Collection of new arXiv papers (2026-05-01) covering systems research for ML training, distributed systems, blockchains, HPC resource management, and embedded real-time systems. Key works: RoundPipe — a round-robin pipeline schedule and transfer/synchronization stack that breaks weight-binding on consumer GPU servers to speed up LLM fine-tuning (1.48–2.16× vs. baselines; enables LoRA on Qwen3-235B at 31K seq on a single 8×RTX4090 server). ZipCCL — a GPU-optimized lossless compression library for collective communication in LLM training, reducing communication time up to 1.35× and end-to-end up