Agrawal et al.: chunked-prefill decode-maximal inference batching

What changed

Split prompt prefills into uniform chunks and batched one compute-intensive chunk with pending token decodes, improving decode utilization and reducing pipeline imbalance.

Scientific scope

Split prompt prefills into uniform chunks and batched one compute-intensive chunk with pending token decodes, improving decode utilization and reducing pipeline imbalance. Original SARATHI is not a claim to all later Sarathi-Serve schedulers. Decode-only and end-to-end speedups differ; workload, GPU, batching and latency constraints control benefit.

Evidence

The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. Original SARATHI is not a claim to all later Sarathi-Serve schedulers. Decode-only and end-to-end speedups differ; workload, GPU, batching and latency constraints control benefit.

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.

• SARATHI: Efficient LLM Inference by Piggybacking Decodes with Chunked Prefills

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