Placement Semantics for Distributed Deep Learning: A Systematic Framework for Analyzing Parallelism Strategies
This provides a unified framework for practitioners to analyze and combine parallelism strategies without trial and error, addressing a bottleneck in training large language models.
The paper tackles the problem of selecting parallelism strategies for distributed deep learning by introducing placement semantics, a systematic framework that predicts memory consumption and communication volume from placement choices alone, matching published results such as ZeRO-3 using 8x less memory than data parallelism at 1.5x communication cost.
Training large language models requires distributing computation across many accelerators, yet practitioners select parallelism strategies (data, tensor, pipeline, ZeRO) through trial and error because no unified systematic framework predicts their behavior. We introduce placement semantics: each strategy is specified by how it places four training states (parameters, optimizer, gradients, activations) across devices using five modes (replicated, sharded, sharded-with-gather, materialized, offloaded). From placement alone, without implementation details, we derive memory consumption and communication volume. Our predictions match published results exactly: ZeRO-3 uses 8x less memory than data parallelism at 1.5x communication cost, as reported in the original paper. We prove two conditions (gradient integrity, state consistency) are necessary and sufficient for distributed training to match single-device results, and provide composition rules for combining strategies safely. The framework unifies ZeRO Stages 1-3, Fully Sharded Data Parallel (FSDP), tensor parallelism, and pipeline parallelism as instances with different placement choices.