DeepSpeed Ulysses: all-to-all sequence and head redistribution

What changed

Sharded tokens across devices and used all-to-all redistribution to give each device full sequence access for a subset of attention heads, followed by inverse redistribution and ZeRO-compatible training.

Scientific scope

Sharded tokens across devices and used all-to-all redistribution to give each device full sequence access for a subset of attention heads, followed by inverse redistribution and ZeRO-compatible training. Inherits sequence/model parallelism, ZeRO and attention kernels. Constant communication volume assumes proportional sequence/device scaling in the paper's model; it is not constant end-to-end runtime. Head-count divisibility/scalability and network topology matter. Original seven-author Microsoft report controls attribution.

Evidence

The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. Inherits sequence/model parallelism, ZeRO and attention kernels. Constant communication volume assumes proportional sequence/device scaling in the paper's model; it is not constant end-to-end runtime. Head-count divisibility/scalability and network topology matter. Original seven-author Microsoft report controls attribution.

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.

• DeepSpeed Ulysses: System Optimizations for Enabling Training of Extreme Long Sequence Transformer 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