large-neural-networks

Introduces PACI, a bubble-free asynchronous pipeline parallel training method that bounds forward/backward weight inconsistency using local gradient accumulation, achieving higher throughput and faster time-to-accuracy without sacrificing stability or memory usage.