Automatic Discovery of Composite SPMD Partitioning Strategies in PartIR
This addresses the challenge of reducing manual effort in optimizing parallelism for large-scale model training, though it appears incremental as it automates an existing process.
The paper tackles the problem of manually identifying efficient composite SPMD partitioning strategies for training large neural networks, and presents an automatic partitioner that matches expert-level strategies for various models.
Large neural network models are commonly trained through a combination of advanced parallelism strategies in a single program, multiple data (SPMD) paradigm. For example, training large transformer models requires combining data, model, and pipeline partitioning; and optimizer sharding techniques. However, identifying efficient combinations for many model architectures and accelerator systems requires significant manual analysis. In this work, we present an automatic partitioner that identifies these combinations through a goal-oriented search. Our key findings are that a Monte Carlo Tree Search-based partitioner leveraging partition-specific compiler analysis directly into the search and guided goals matches expert-level strategies for various models.