{ "schema_version": 1, "builder_plan": { "scenario": "Profiling shows attention computation is a bottleneck in your self-hosted service.", "change": "Check the supported GPU, tensor shapes, and precision, then replace only the relevant kernel.", "baseline": "The existing kernel inside the same complete serving pipeline.", "steps": [ "Locate the paper's implementation and check its license, model support, data needs, and hardware requirements. Stop here if you cannot obtain the required artifacts.", "Save a small set of representative product tasks and expected outcomes. Keep evaluation tasks out of any training or tuning, and choose quality and cost limits before running.", "Run the baseline and the proposed change on identical inputs. Keep other settings fixed and repeat paired runs to expose variation.", "Save per-task outputs, failures, timings, and all usage costs. Review regressions before expanding the test." ], "measure": "Numerical differences, end-to-end latency, throughput, and memory at realistic shapes.", "decision": "Keep the change only if it meets your preselected quality and latency limits and improves the metric you care about after setup and operating costs. Otherwise retain the baseline. This is an exploration plan, not a validated recommendation." }, "builder_review": null, "arxiv_id": "2609.35629", "title": "SANTA++: Sampling Attention through Representative Keys", "summary": "SANTA++ replaces full dense attention over the KV cache with a training-free importance-sampled estimator that scores one representative key per team, samples a fixed budget of teams, and computes exact attention only within them. It costs an unbiased-estimation variance tradeoff: accuracy drops to 85%-91% of baseline on RULER, and realized speedup depends on the custom kernel rather than stock FlashAttention. It can fail on workloads where relevant tokens scatter across many teams or where the sampling budget is set too low, and gains beyond 32K contexts or on models other than the tested one are unvalidated.", "rationale": "The method is training-free, open-source, and directly reduces KV reads and attention latency on a single GPU, but the quality variance and kernel dependency require validation on the reader's specific long-context workload before production adoption.", "technique": "Stochastic representative-key attention sampling", "family": "kernels_e_atencao", "recommendation": "testar", "published": "2026-09-28", "as_of": "2026-10-18", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2609.35629", "independent_implementations": 0, "total_implementations": 0, "stars": 0, "citations": null, "claims": [ { "claim_id": "claim-01", "statement": "SANTA++ replaces full dense attention over the KV cache with a training-free importance-sampled estimator that scores one representative key per team, samples a fixed budget of teams, and computes exact attention only within them. It costs an unbiased-estimation variance tradeoff: accuracy drops to 85%-91% of baseline on RULER, and realized speedup depends on the custom kernel rather than stock FlashAttention. It can fail on workloads where relevant tokens scatter across many teams or where the sampling budget is set too low, and gains beyond 32K contexts or on models other than the tested one are unvalidated.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2609.35629", "source_page": null, "source_excerpt": "" }, { "claim_id": "claim-02", "statement": "With 32-64 sampled teams, SANTA++ reads 16%-22% of dense attention's KV entries while retaining 94%-99% of baseline scores on LongBench v2 and HELMET RAG (85%-91% on RULER) with Qwen2.5-7B-Instruct at 32K context, and its kernel delivers a 1.69x attention speedup over FlashAttention at 32K with 31 teams.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2609.35629", "source_page": null, "source_excerpt": "" } ], "exposure_map": [ { "dimension": "quality", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "compute", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "latency", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "operations", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "compatibility", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "security", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "data_and_training", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "reproducibility", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" } ], "risks": [], "minimum_test": [], "open_questions": [], "independent_tests": [], "lab_reproductions": [], "report_available": false, "validation_tier": "unknown", "evidence_tier": "unknown", "infrastructure_basis": "unknown", "equations": [], "equations_status": "not_fetched", "core_kind": "", "equations_fetched_at": "", "equations_source_url": "", "equations_source_sha256": "", "exposures_reviewed_at": "", "exposures_source_sha256": "" }