HeLoCo: Efficient asynchronous low-communication training under data and device heterogeneity

Collection of systems and algorithms for efficient, robust, and energy-aware ML and multi-agent computation under hardware and data heterogeneity. Key contributions include HeLoCo (direction-aware correction for asynchronous low-communication training to mitigate stale/misaligned pseudo-gradients in non-IID and heterogeneous settings); Augur (pre-execution energy prediction for scientific workflows enabling energy/carbon-aware scheduling); the Cartan-Topos Protocol (geometric/categorical framework for resilient multi-agent coordination using manifolds, sheaves, and topos-based temporal logic);