Rajbhandari and colleagues: ZeRO distributed model-state partitioning

What changed

Partitioned optimizer states, gradients and model parameters across data-parallel workers to reduce replicated training memory.

Scientific scope

Partitioned optimizer states, gradients and model parameters across data-parallel workers to reduce replicated training memory. Communication and memory tradeoffs depend on the ZeRO stage. The title's trillion-parameter target is not itself an experimental training result.

Evidence

The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. Communication and memory tradeoffs depend on the ZeRO stage. The title's trillion-parameter target is not itself an experimental training result.

Contributors

• Source-listed authors and technical contributors · SCIENCE: Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.

• ZeRO: Memory Optimizations Toward Training Trillion Parameter Models

Provisional WRS assessment, edition 2026-10-16.1. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.

Full scientific record · Browse contribution grades · Request a correction