{ "schema_version": 1, "builder_plan": { "scenario": "Long customer conversations exhaust memory or limit concurrent requests.", "change": "Change only the attention cache policy, after checking model and runtime support.", "baseline": "The same model with its ordinary attention cache.", "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": "Peak cache memory, long-context answer quality, first-token latency, and requests completed.", "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": "2610.28385", "title": "RACE: RoPE-Aligned Attention Compression Based on Query-Key Activation Energies", "summary": "RACE replaces full-dimensional per-head keys and queries (and reconstruction-based KV-cache compression) with a natively reduced-dimensional attention formed by selecting complete RoPE rotary pairs ranked by joint query-key activation energy, folding the selection into the model weights post-training. It costs a one-time calibration pass and permanently removes up to half of the head dimension, trading some representational capacity for lower KV-cache memory and attention FLOPs without custom kernels. It can fail if the calibration set is unrepresentative of production inputs, if accuracy on zero-shot or long-context tasks degrades at the chosen compression level, or if short-context workloads see little benefit since the gains concentrate at long contexts.", "rationale": "It is a post-training, kernel-free method compatible with a single-GPU budget and offers meaningful memory and throughput gains, but the 1.79x figure is measured at 256K context and quality retention depends on calibration, so workload-specific validation is required before production adoption.", "technique": "RoPE-aligned head-dimension KV-cache compression via rotary-pair selection", "family": "cache_kv", "recommendation": "testar", "published": "2026-10-20", "as_of": "2026-10-21", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2610.28385", "independent_implementations": 0, "total_implementations": 1, "stars": 122, "citations": null, "claims": [ { "claim_id": "claim-01", "statement": "RACE replaces full-dimensional per-head keys and queries (and reconstruction-based KV-cache compression) with a natively reduced-dimensional attention formed by selecting complete RoPE rotary pairs ranked by joint query-key activation energy, folding the selection into the model weights post-training. It costs a one-time calibration pass and permanently removes up to half of the head dimension, trading some representational capacity for lower KV-cache memory and attention FLOPs without custom kernels. It can fail if the calibration set is unrepresentative of production inputs, if accuracy on zero-shot or long-context tasks degrades at the chosen compression level, or if short-context workloads see little benefit since the gains concentrate at long contexts.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2610.28385", "source_page": null, "source_excerpt": "" }, { "claim_id": "claim-02", "statement": "At 50% KV-cache retention and 256K context, RACE reduces total inference FLOPs by 47.8%, yielding 1.79x higher throughput, 43.2% lower TTFT, and 28.1% lower allocated GPU memory versus the uncompressed model.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2610.28385", "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": "" }