1 paper
Arthur Douillard, Keith Rush, Yani Donchev +14
Modern large-scale language model pre-training relies heavily on the single program multiple data (SPMD) paradigm, which requires tight coupling across accelerators. Due to this co…