Accelerating large-scale distributed neural network training with SPMD parallelism