CUCo: An Agentic Framework for Compute and Communication Co-design

This collection of arXiv announcements covers advances in GPU and distributed-systems tooling, correctness analysis, and empirical security assessments. Notable items: CUCo — an agentic, training-free workflow that auto-generates CUDA kernels co-optimizing computation and communication to reduce end-to-end latency (up to 1.57x); a formal treatment of the “Forward-In-Time-Only” (FITO) category mistake proving that atomic checkpoints and atomic firmware deployments are unattainable under usual asynchronous/crash-recovery assumptions and proposing bilateral convergence protocols; an empirical, 비교