AWhat changed
Combined composable block-sparse KV-cache representations, JIT-customizable attention templates and dynamically load-balanced scheduling compatible with static CUDA graphs.
Scientific scope
Combined composable block-sparse KV-cache representations, JIT-customizable attention templates and dynamically load-balanced scheduling compatible with static CUDA graphs. FlashAttention, paged KV storage and radix caching are inherited. This is serving-kernel integration/customization, not an invention of exact attention or a universal latency reduction; source benchmarks are not independently replayed.
Evidence
The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. FlashAttention, paged KV storage and radix caching are inherited. This is serving-kernel integration/customization, not an invention of exact attention or a universal latency reduction; source benchmarks are not independently replayed.
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.
- FlashInfer: Efficient and Customizable Attention Engine for LLM Inference Serving
- Ye et al.: customizable load-balanced attention kernels for serving — original project
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