AWhat changed
Partitioned model training between GPU execution and CPU optimizer state/computation, combining communication-aware placement, a CPU optimizer and delayed updates with ZeRO data parallelism.
Scientific scope
Partitioned model training between GPU execution and CPU optimizer state/computation, combining communication-aware placement, a CPU optimizer and delayed updates with ZeRO data parallelism. Depends on CPU compute/memory, transfer bandwidth, batch size and specific delayed-update choices. Reported scale and performance are hardware-specific; original ZeRO state sharding remains a distinct inherited contribution.
Evidence
The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. Depends on CPU compute/memory, transfer bandwidth, batch size and specific delayed-update choices. Reported scale and performance are hardware-specific; original ZeRO state sharding remains a distinct inherited contribution.
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-Offload: Democratizing Billion-Scale Model Training
- Ren et al.: ZeRO-Offload heterogeneous training — author/proceedings record
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