{ "schema_version": 1, "builder_plan": { "scenario": "Your self-hosted model spends too long generating answers token by token.", "change": "Check whether a compatible draft model and serving integration exist before renting compute.", "baseline": "The same target with speculation off, then your existing drafter if you have one.", "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": "Accepted draft length, tokens per second, tail latency, output quality, and total GPU cost.", "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": "2608.30252", "title": "Strong Drafts Need Compact Memories: Long-Context Speculative Decoding with Compressed KV Cache", "summary": "The method replaces the draft model's full KV cache in long-context speculative decoding with a compact memory built and incrementally updated by a lightweight adaptor, while the target model keeps its full cache and standard verification. It costs an additional adaptor component to train or obtain and integrate, extra implementation complexity in the draft path, and draft-side memory that must be refreshed alongside generation; the paper does not state adaptor training cost. It can fail if acceptance rates drop on workloads whose long-range dependencies the compressed memory does not retain, or if adaptor inference overhead erodes the per-step latency advantage at shorter prefixes.", "rationale": "The speedup claim is substantial and lossless, but adoption requires a compatible memory-augmented draft (not an off-the-shelf artifact) and workload-specific validation of acceptance rates, so a constrained team should pilot it on their own long-context traffic before committing.", "technique": "Speculative decoding with compressed draft-side KV memory", "family": "decodificacao_especulativa", "recommendation": "testar", "published": "2026-08-31", "as_of": "2026-10-18", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2608.30252", "independent_implementations": 0, "total_implementations": 0, "stars": 0, "citations": 0, "claims": [ { "claim_id": "claim-01", "statement": "The method replaces the draft model's full KV cache in long-context speculative decoding with a compact memory built and incrementally updated by a lightweight adaptor, while the target model keeps its full cache and standard verification. It costs an additional adaptor component to train or obtain and integrate, extra implementation complexity in the draft path, and draft-side memory that must be refreshed alongside generation; the paper does not state adaptor training cost. It can fail if acceptance rates drop on workloads whose long-range dependencies the compressed memory does not retain, or if adaptor inference overhead erodes the per-step latency advantage at shorter prefixes.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2608.30252", "source_page": null, "source_excerpt": "" }, { "claim_id": "claim-02", "statement": "Up to 2.08x (Llama 3.1-8B) and 3.33x (70B) speedup over autoregressive decoding at prefixes up to 32K, with over 70% reduction in draft-side KV memory and lossless output guarantees.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2608.30252", "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": [ { "anchor": "S3.E1", "label": "(1)", "section": "3 Background & Motivation", "role": "baseline", "latex": "\\text{Speedup}=\\frac{L_{\\mathrm{acc}}+1}{L_{\\text{draft}}\\cdot\\tfrac{t_{d}}{t_{t}}+1}.", "mathml": "Speedup=Lacc+1Ldraft⋅tdtt+1.", "context": "Speculative decoding accelerates autoregressive generation through a draft–verify paradigm that preserves the target model’s output distribution (Leviathan et al., 2023). Given a target model 𝖬t and a cheaper draft model 𝖬d, each iteration first runs 𝖬d for Ldraft autoregressive steps to propose a candidate continuation, and then invokes 𝖬t in a single parallel forward pass to verify the candidates. Let Lacc denote the mean number of accepted draft tokens per iteration. Standard SD emits these tokens plus one target token: a correction token after the first rejection or a bonus token when all candidates are accepted. Thus, ignoring early termination, the mean number of emitted tokens is Lemit=Lacc+1. Let td and tt denote the single-token latency of 𝖬d and 𝖬t, respectively. The decoding speedup over standard decoding is", "headline": "", "explanation": "", "objective": "", "symbols": {}, "evidence": "", "evidence_url": "", "source_url": "" }, { "anchor": "S3.E2", "label": "(2)", "section": "3 Background & Motivation", "role": "complexity", "latex": "T_{\\text{decode}}\\;\\approx\\;\\alpha\\,d^{2}\\;+\\;\\beta\\,L\\,d,", "mathml": "Tdecode≈α​d2+β​L​d,", "context": "A stronger independent draft model recovers accepted length, but full-KV drafting requires reading the entire historical KV cache at every draft step. We decompose single-token decode latency into a weight-related term and a KV-access term:", "headline": "", "explanation": "", "objective": "", "symbols": {}, "evidence": "", "evidence_url": "", "source_url": "" }, { "anchor": "S4.E3", "label": "(3)", "section": "4 Method", "role": "proposed_method", "latex": "\\mathcal{M}_{t}=\\mathcal{M}_{\\mathrm{sink}}\\cup\\mathcal{M}_{\\mathrm{local},t}\\cup\\mathcal{M}_{\\mathrm{slot},t},", "mathml": "ℳt=ℳsink∪ℳlocal,t∪ℳslot,t,", "context": "We augment the window with memory slots: compact states materialized at a fixed rate from completed history and kept in the draft cache after their raw KV is released. At step t, the draft-side working memory is", "headline": "", "explanation": "", "objective": "", "symbols": {}, "evidence": "", "evidence_url": "", "source_url": "" } ], "equations_status": "selected", "core_kind": "formula", "equations_fetched_at": "2026-09-23", "equations_source_url": "", "equations_source_sha256": "", "exposures_reviewed_at": "", "exposures_source_sha256": "" }