What? It definitely is. Data parallelism, model parallelism, parameter server to workers, MoE itself can be split up, etc. But even if it wasn’t, you can simp

What? It definitely is. Data parallelism, model parallelism, parameter server to workers, MoE itself can be split up, etc. But even if it wasn’t, you can simply parallelize training runs with slight variations in hyperparameters. That is what the article is describing.