{"paper":{"title":"FlashInfer: Efficient and Customizable Attention Engine for LLM Inference Serving","license":"http://creativecommons.org/licenses/by-sa/4.0/","headline":"FlashInfer uses block-sparse KV-cache formats and JIT-compiled attention templates to cut inter-token latency by 29-69% in LLM serving.","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.DC","authors_text":"Arvind Krishnamurthy, Baris Kasikci, Lequn Chen, Luis Ceze, Ruihang Lai, Stephanie Wang, Tianqi Chen, Vinod Grover, Wuwei Lin, Yineng Zhang, Zihao Ye","submitted_at":"2025-01-02T02:02:20Z","abstract_excerpt":"Transformers, driven by attention mechanisms, form the foundation of large language models (LLMs). As these models scale up, efficient GPU attention kernels become essential for high-throughput and low-latency inference. Diverse LLM applications demand flexible and high-performance attention solutions. We present FlashInfer: a customizable and efficient attention engine for LLM serving. FlashInfer tackles KV-cache storage heterogeneity using block-sparse format and composable formats to optimize memory access and reduce redundancy. It also offers a customizable attention template, enabling ada"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Compared to state-of-the-art LLM serving solutions, FlashInfer achieve 29-69% inter-token-latency reduction compared to compiler backends for LLM serving benchmark, 28-30% latency reduction for long-context inference, and 13-17% speedup for LLM serving with parallel generation.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The reported speedups assume that the block-sparse format and JIT templates integrate cleanly with existing serving frameworks without hidden overheads from compilation or scheduling that would appear under production load patterns not tested in the benchmarks.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"FlashInfer delivers a customizable attention kernel that reduces inter-token latency by 29-69% in LLM serving benchmarks via optimized KV-cache storage and load-balanced scheduling compatible with CUDA graphs.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"FlashInfer uses block-sparse KV-cache formats and JIT-compiled attention templates to cut inter-token latency by 29-69% in LLM serving.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"b4b44bab60ff148d0db56ba6d96d2b1a261dac13863d61f6f6848b3b77d3d25b"},"source":{"id":"2501.01005","kind":"arxiv","version":2},"verdict":{"id":"d2941e40-5d8c-4a82-b667-2b1836ce3227","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T13:21:57.766174Z","strongest_claim":"Compared to state-of-the-art LLM serving solutions, FlashInfer achieve 29-69% inter-token-latency reduction compared to compiler backends for LLM serving benchmark, 28-30% latency reduction for long-context inference, and 13-17% speedup for LLM serving with parallel generation.","one_line_summary":"FlashInfer delivers a customizable attention kernel that reduces inter-token latency by 29-69% in LLM serving benchmarks via optimized KV-cache storage and load-balanced scheduling compatible with CUDA graphs.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The reported speedups assume that the block-sparse format and JIT templates integrate cleanly with existing serving frameworks without hidden overheads from compilation or scheduling that would appear under production load patterns not tested in the benchmarks.","pith_extraction_headline":"FlashInfer uses block-sparse KV-cache formats and JIT-compiled attention templates to cut inter-token latency by 29-69% in LLM serving."},"references":{"count":14,"sample":[{"doi":"10.1145/1583991.1584053","year":2004,"title":"URL https://arxiv.org/abs/2004. 05150. Buluç, A., Fineman, J. T., Frigo, M., Gilbert, J. R., and Leiserson, C. E. Parallel sparse matrix-vector and matrix-transpose-vector multiplication using com- pr","work_id":"4e1afef5-52c1-4ec8-8c63-7c4bf85ce428","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.1145/3458817.3476182","year":2024,"title":"URL https: //doi.org/10.1145/3458817.3476182","work_id":"f104f07e-0962-463f-959e-ddecd7f1f7ba","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.1177/1094342004041296","year":2024,"title":"org/paper_files/paper/2024/file/ 5321b1dabcd2be188d796c21b733e8c7- Paper-Conference.pdf","work_id":"db04265b-d798-47c9-b726-4ca1021628a1","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.1109/sc41404.2022.00042","year":2022,"title":"El-rec: Efficient large-scale recommendation model training via tensor-train embedding table","work_id":"0625ddd0-436d-4c7a-b66a-97409cfc830c","ref_index":7,"cited_arxiv_id":"1404.2022","is_internal_anchor":true},{"doi":"10.1145/55364.55378","year":2024,"title":"URL https://doi","work_id":"a5dd3d03-6672-4aae-acf0-c19739d4a596","ref_index":9,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":14,"snapshot_sha256":"d5851aafbceae1d6888f616320a215806f2f3d523cfea2ed52d3d251580d9b49","internal_anchors":3},"formal_canon":{"evidence_count":2,"snapshot_sha256":"1d4b34b6a74bdcbf05457dc3fcba8230a664874232639143fd07d4e35971cce2"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}