Model Parallelism With Subnetwork Data Parallelism

arXiv:2507.09029v5 Announce Type: replace-cross Abstract: Pre-training large neural networks at scale imposes heavy memory demands on accelerators and often requires costly communication. We introduce Subnetwork Data Parallelism (SDP), a distributed training framework that partitions a model into structured subnetworks trained across workers without exchanging activations. We study two complementary masking regimes: backward masking, which applies sparsity only in the backward step to retain unb