{"id":"arxiv-2606.30389v1","name":"predict-reuse-and-repair-accelerating-dynamic-sparse-attention-for-long","tags":["arxiv-import","external-catalog","arxiv","paper","cs.LG"],"task":"evaluation","type":"collection","evals":[],"owner":"arxiv","stars":0,"formats":["atom","pdf","paper"],"hosting":{"status":"blocked","storageProvider":"local","hostedFiles":0,"totalFiles":0,"hostedFilesAuthority":"live-serve-decision","hosted":false,"downloadable":false,"readyFileCount":0,"serveAuthority":"live-serve-decision","completeRunnable":false,"authority":"postgres-executable-state-snapshot","authorityDigest":null,"downloadReady":false,"mirroredFiles":0,"mirroredBytes":0,"runReady":false,"verified":false},"license":"unknown","summary":"Predict, Reuse, and Repair: Accelerating Dynamic Sparse Attention for Long-Context LLM Decoding","version":"2606.30389v1","modality":"paper","upstream":{"doi":"","gated":false,"repoId":"2606.30389v1","authors":["Tianyu Wang","Gourav Rattihalli","Aditya Dhakal","Junbo Li","Zhiwei Ren","Dejan Milojicic","Longfei Shangguan"],"private":false,"repoUrl":"http://arxiv.org/abs/2606.30389v1","disabled":false,"provider":"arxiv","repoType":"paper","revision":"2606.30389v1","fetchedAt":"2026-07-01T00:12:49.224Z","categories":["cs.LG"],"importMode":"external-catalog","licenseRaw":"arxiv-metadata"},"versions":[{"commit":"2606.30389v1","version":"2606.30389v1","releasedAt":"2026-06-29T14:43:25Z","scanStatus":"pending","manifestDigest":"external-not-mirrored"}],"citations":0,"createdAt":"2026-06-29T14:43:25Z","framework":"arxiv","publisher":"Tianyu Wang","sizeBytes":1,"sourceUrl":"http://arxiv.org/abs/2606.30389v1","updatedAt":"2026-06-29T14:43:25Z","scanStatus":"pending","visibility":"public","description":"External arXiv paper metadata for 2606.30389v1. Dynamic sparse attention (DSA) accelerates long-context LLM decoding by attending to only the top-K KV blocks relevant to each query, but it introduces a serialized selection-to-attention dependency that emerges as a new latency bottleneck. We present PRR, a speculate-reuse-repair runtime that exploits temporal locality in DSA selections to predict likely blocks, speculate the attention over them while selection is in flight, and incrementally repair missed blocks once the true selected set is known. PRR uses a lightweight EMA-based predictor, a profiling-guided speculation budget that keeps speculative work off the critical path, and a FlashAttention-based repair kernel that folds missed blocks into the partial attention state using online-softmax statistics. Across long-context benchmarks and representative DSA methods, PRR reduces per-token decoding latency by up to 40% while preserving downstream task accuracy. Github: https://github.com/Tianyu9748/Incremental_FlashAttention Hugging Bay stores this as paper/context metadata only; full text and source files stay upstream.","mirrorCount":0,"safetyNotes":["External metadata imported from arxiv on 2026-07-01T00:12:49.224Z.","This record is searchable and linkable, but Hugging Bay does not host its files.","Use the upstream source URL until a source-specific legal and mirror review approves local hosting."],"releaseGates":{"scan":"pending","files":"pending","license":"unknown","metadata":"passed","signature":"pending","provenance":"passed"},"downloadCount":0,"cardCompleteness":92,"verificationStatus":"pending","requestedVisibility":"public","licenseLabel":"arxiv-metadata","repo":"arxiv/predict-reuse-and-repair-accelerating-dynamic-sparse-attention-for-long","files":[],"fileCount":0,"hostedFileCount":0,"externalFileCount":0,"communityReviewSummary":{"count":0,"averageRating":0,"verifiedRuns":0,"selfReportedVerifiedRuns":0,"verificationQualification":"quarantined-untrusted-evidence","verificationTrustImpact":"none","helpfulCount":0,"reportCount":0,"reportedReviews":0,"distribution":{"1":0,"2":0,"3":0,"4":0,"5":0},"toolCounts":{},"categoryAverages":{},"categoryCounts":{},"benchmarkEvidence":{"count":0,"verifiedCount":0,"benchmarkCount":0,"performanceCount":0,"hardwareCounts":{},"topMetrics":[],"recent":[]},"recent":[]},"publicDownloadableFileCount":0,"publicFileCount":0,"displayOnly":true,"authorizationRequired":true,"eligibilitySnapshotAt":"2026-11-22T01:08:04.401Z","webUrl":"https://huggingbay.xyz/artifact/arxiv-2606.30389v1","listedFileLowerBound":0,"hostingStatus":"blocked","signals":{"size":{"known":false,"bytes":0,"label":"unknown","reason":"upstream file sizes are not indexed yet"},"demand":{"mirrorRequests":0,"upstreamDownloads":0,"upstreamStars":0,"label":"no demand yet","basis":"none"},"trust":{"score":0,"level":"unknown","label":"Unknown","qualification":"recorded-state-unknown","trustImpact":"none","checks":[{"id":"license","label":"License","status":"fail","detail":"arxiv-metadata"},{"id":"provenance","label":"Provenance","status":"pass","detail":"2606.30389v1"},{"id":"scan","label":"Scan","status":"warn","detail":"pending"},{"id":"hosting","label":"Hosted files","status":"warn","detail":"external metadata only"},{"id":"manifest","label":"Byte manifest integrity","status":"warn","detail":"not final yet"}],"strengths":["Provenance: 2606.30389v1"],"gaps":["License: arxiv-metadata","Scan: pending","Hosted files: external metadata only","Byte manifest integrity: not final yet"]},"readiness":{"score":0,"level":"unknown","hostedFinal":false,"hostedFiles":0,"externalFiles":0,"totalFiles":0},"action":{"id":"open-source","label":"Open official source","detail":"Use the official immutable source while Hugging Bay evidence remains incomplete.","href":"http://arxiv.org/abs/2606.30389v1"}},"mirrorCompletion":{"status":"external","hostedFiles":0,"totalFiles":0,"externalFiles":0,"hostedBytes":0,"externalBytes":1,"lowRiskExternalFiles":0,"overFileCapFiles":0,"overBatchCapFiles":0,"directMirrorExternalFiles":0,"directMirrorExternalBytes":0,"currentLimits":{"maxMirrorFileBytes":536870912,"maxMirrorTotalBytes":1073741824,"maxMirrorFiles":8,"maxDirectMirrorFileBytes":8589934592,"maxDirectMirrorTotalBytes":17179869184},"directMirrorExternalSamples":[],"largestExternalFiles":[]},"recommendedFile":null,"torrent":null,"torrentUrl":null,"magnetUri":null,"downloadUrl":null,"landingUrl":null,"manifestUrl":null,"localKitUrl":null,"licenseClarity":{"license":"arxiv-metadata","commercialUse":"unknown","confidence":"none","obligations":[],"summary":"No license stated","redistribution":"Check the upstream license text or contact the model author before deciding how to use it.","notes":"Not legal advice."},"licenseObligationsDigest":null,"licenseUpstreamRecord":{"declaredLicense":"arxiv-metadata","upstreamLicenseRaw":"arxiv-metadata","sourceUrl":"http://arxiv.org/abs/2606.30389v1","upstream":{"doi":"","gated":false,"repoId":"2606.30389v1","authors":["Tianyu Wang","Gourav Rattihalli","Aditya Dhakal","Junbo Li","Zhiwei Ren","Dejan Milojicic","Longfei Shangguan"],"private":false,"repoUrl":"http://arxiv.org/abs/2606.30389v1","disabled":false,"provider":"arxiv","repoType":"paper","revision":"2606.30389v1","fetchedAt":"2026-07-01T00:12:49.224Z","categories":["cs.LG"],"importMode":"external-catalog","licenseRaw":"arxiv-metadata"},"licenseLabel":"arxiv-metadata"},"gateStatus":"ungated","downloadable":false,"downloadStatus":"BLOCKED","reasonCodes":["incomplete-manifest","no-tested-run-plan","unverified-bytes"],"reason":"Files withheld — not faked. This model is listed, but we haven't confirmed its download files, so no download URLs are published. What's missing: file manifest — not final; downloadable files — 0 confirmed; run plan — not tested.","serveEvaluatedAt":"2026-11-22T01:09:36.905Z","serveEvaluationId":"serve-c2bd7d2ff766c16400586466fe91823f23e9914e4f23f390ebb380805d83fefc","readiness":"blocked","downloadReady":false,"mirroredBytes":0,"readinessCheckedAt":"2026-11-22T01:09:36.974Z","readinessSource":"file-check","serveExplanation":{"downloadable":false,"downloadStatus":"BLOCKED","reasonCodes":["incomplete-manifest","no-tested-run-plan","unverified-bytes"],"reason":"Files withheld — not faked. This model is listed, but we haven't confirmed its download files, so no download URLs are published. What's missing: file manifest — not final; downloadable files — 0 confirmed; run plan — not tested.","serveEvaluatedAt":"2026-11-22T01:09:36.905Z","serveEvaluationId":"serve-c2bd7d2ff766c16400586466fe91823f23e9914e4f23f390ebb380805d83fefc","readiness":"blocked","downloadReady":false,"hostedFileCount":0,"mirroredBytes":0,"readinessCheckedAt":"2026-11-22T01:09:36.905Z","readinessSource":"file-check","degraded":false,"stale":false},"licenseState":"unknown","licenseUnresolved":true,"licenseCategory":"no-license","commandsAvailable":false,"runReady":false,"attribution":"Cataloged open AI model · https://huggingbay.xyz/artifact/arxiv-2606.30389v1","decentralizedDistribution":{"protocol":"bittorrent-magnet-metadata","verifiedFallbacks":0,"pendingReviewRequests":0,"reviewRequired":true,"policy":"Peer-assisted fallbacks are metadata only until Hugging Bay review confirms source, license, provenance, and safety. Gated/private bypasses are not allowed."},"hosted":false,"modelState":{"state":"not_hosted","word":"Not hosted here","sentence":"Files are not hosted here yet. Open the model to ask for them."},"pageFacts":{"id":"arxiv-2606.30389v1","repo":"arxiv/predict-reuse-and-repair-accelerating-dynamic-sparse-attention-for-long","type":"collection","purpose":"collection artifact for evaluation","license":{"id":"unknown","label":"arxiv-metadata","state":"unknown","declared":"arxiv-metadata","url":"","source":"unconfirmed","hostedFile":false,"commercialUse":"unknown","terms":"arxiv-metadata"},"task":{"id":"evaluation","label":"evaluation"},"size":{"bytes":0,"label":"unknown","known":false,"repositoryLabel":"Full repo upstream","hostedKnown":false,"hostedBytes":0,"hostedLabel":"size unknown","hostedSummary":"availability unknown","hostedIsSubset":false,"downloadBytes":0,"downloadLabel":"","recommended":null,"parameters":null,"sentence":"Hosted here: availability unknown. The full repository size is not recorded."},"files":{"hosted":0,"total":0,"label":"0 hosted of 0 listed"},"download":{"status":"blocked","gated":"serve","label":"Downloadable files: 0 confirmed","url":""},"released":{"iso":"2026-06-29","label":"Jun 29, 2026","upstream":true,"factLabel":"Released","licenseLabel":"arxiv-metadata"},"updated":{"iso":"2026-06-29","label":"Jun 29, 2026","factLabel":"Updated at source"},"source":{"provider":"arxiv","label":"arxiv","url":"","repoId":"2606.30389v1"},"languages":[],"baseModel":{"ref":"","url":"","resolveUrl":""},"paperIds":[],"arxivIds":[],"sameAs":[],"creator":{"@type":"Organization","name":"Tianyu Wang"},"scholarlyArticles":[],"scanStatus":"pending","faq":[{"id":"what-is-it","question":"What is arxiv/predict-reuse-and-repair-accelerating-dynamic-sparse-attention-for-long?","answer":"arxiv/predict-reuse-and-repair-accelerating-dynamic-sparse-attention-for-long is a collection artifact for evaluation published by Tianyu Wang, released Jun 29, 2026. Predict, Reuse, and Repair: Accelerating Dynamic Sparse Attention for Long-Context LLM Decoding"},{"id":"license","question":"Can I use it commercially?","answer":"No license stated."},{"id":"size","question":"How big is arxiv/predict-reuse-and-repair-accelerating-dynamic-sparse-attention-for-long?","answer":"Hosted here: availability unknown. The full repository size is not recorded."},{"id":"download","question":"Where can I download arxiv/predict-reuse-and-repair-accelerating-dynamic-sparse-attention-for-long?","answer":"This model is listed, but we haven't confirmed its download files, so we won't offer a download button we can't stand behind. The record lists 0 files."},{"id":"safety","question":"Has arxiv/predict-reuse-and-repair-accelerating-dynamic-sparse-attention-for-long been scanned for unsafe files?","answer":"Source scan status: pending. A scan checks the files for known-bad signatures; it does not mean the model is safe. The machine-readable verdict is at /api/artifacts/arxiv-2606.30389v1/safety on Hugging Bay."},{"id":"fit-unknown","question":"Will arxiv/predict-reuse-and-repair-accelerating-dynamic-sparse-attention-for-long run on my computer?","answer":"We can’t check the fit; we don’t know your machine. We don’t have your machine’s specs, so no check ran and no verdict is shown. Answer two quick questions and we’ll estimate, from the memory you tell us, whether it fits your machine. Or skip it; you can still download; fit isn’t checked."}]},"guidedDownloadCard":{"schemaVersion":"hugging-bay.guided-download-card/v2","artifactId":"arxiv-2606.30389v1","repo":"arxiv/predict-reuse-and-repair-accelerating-dynamic-sparse-attention-for-long","serveState":"BLOCKED","downloadReady":false,"reasonCodes":["manifest_not_final","files_unconfirmed","run_plan_untested"],"withheld":{"active":true,"reasonCodes":["manifest_not_final","files_unconfirmed","run_plan_untested"]},"wasmRun":null,"bytesOnStorage":false,"file":null,"noHostedWeights":true,"notSingleFile":false,"download":{"available":false,"url":null},"otherFiles":[],"run":{"available":false,"reason":"no-runnable-file","lmStudioUrl":"https://lmstudio.ai/","terminalCommand":null,"verificationMessage":null},"paused":{"active":false},"blocked":{"active":true,"reason":"page-facts-not-hosted","message":"Files withheld. Not faked."},"unconfirmed":{"active":false},"licenseGate":{"active":false},"checks":{"license":null,"licenseLabel":"arxiv-metadata","fingerprintMatches":false,"hostedFiles":0,"totalFiles":0},"policyVersion":"hugging-bay.guided-download-card/v2","evaluatedAt":"2026-11-22T01:09:36.977Z","readiness":"blocked","downloadable":false,"downloadStatus":"BLOCKED","hostedFileCount":0,"mirroredBytes":0,"readinessCheckedAt":"2026-11-22T01:09:36.977Z","readinessSource":"file-check","degraded":false,"stale":false,"hosted":false,"hostingStatus":"blocked","publicDownloadableFileCount":0,"licenseLabel":"arxiv-metadata","serve_verdict":{"can_download":false,"bytes_verified":false,"signature":"unsigned","freshness":"current","checked_at":"2026-11-22T01:09:36.977Z","evidence_url":"/api/manifests/arxiv-2606.30389v1"}},"howToRun":{"schema":"hugging-bay.how-to-run.v1","available":false,"runtime":null,"command":null,"file":null,"minimumRamGb":null,"downloadUrl":null,"downloadCommand":null,"verifyCommand":null,"downloadUrlNote":"Stable URL. Each GET re-checks the file and redirects to a signed link that lasts at most 300 seconds; open it again for a fresh link.","filesUrl":"https://huggingbay.xyz/api/v1/artifacts/arxiv-2606.30389v1/files","tested":false,"note":"No single hosted file is recorded for this model yet. minimumRamGb is an estimate from the file size (file at most 0.6 x memory), not a measurement."},"canIRun":true,"runTested":false,"degraded":false,"stale":false,"contentControls":{"state":"not assessed","basis":"no_explicit_uploader_evidence","matchedTerms":[],"assessedAt":null},"contentControlsLabel":"Content controls: not assessed.","contentControlsAttribution":{"upstream":{"provider":"arxiv","repoId":"2606.30389v1"},"licenseLabel":"arxiv-metadata"},"runs_here":{"schema_version":1,"artifact_id":"arxiv-2606.30389v1","records":[],"next_cursor":null},"hardware_fit":{"schema":"hugging-bay.hardware-fit.v1","artifactId":"arxiv-2606.30389v1","revision":"","claim":"estimated_physical_memory_capacity","runnableAuthority":false,"representations":[],"reason":"immutable_revision_unavailable","licenseLabel":"arxiv-metadata"},"note":{"summary":"Downloadable doesn't mean tested: this record says whether the exact file can be downloaded now, not that the model is complete or that it runs.","downloadable":false},"serve_verdict":{"can_download":false,"bytes_verified":false,"signature":"unsigned","freshness":"current","checked_at":"2026-11-22T01:09:36.974Z","evidence_url":"/api/manifests/arxiv-2606.30389v1"}}