{ "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": "2607.27269", "title": "Beyond KV Reconstruction: Functional Reconstruction for MLA Draft Models in Speculative Decoding", "summary": "The method replaces direct low-rank MHA/GQA-to-MLA conversion for draft models with a post-conversion calibration procedure that fine-tunes each converted MLA attention module to reproduce the original module's post-output-projection output on calibration hidden states, without verifier logits or supervision. Its cost is an additional per-module optimization stage after conversion, while the converted MLA cache layout and inference graph are preserved, so no extra serving-time memory or latency is introduced. It can fail when acceptance gains do not materialize (26 of 64 cells unchanged, one degraded), when calibration hidden states are unrepresentative of the production workload, or when the reader's stack uses MHA/GQA drafts directly and the MLA conversion step itself is not warranted.", "rationale": "The technique addresses a real acceptance-rate regression in converted MLA drafts and is feasible on a single GPU, but the gain is inconsistent across configurations, so the reader should validate it on their specific draft-target pair and backend before production adoption.", "technique": "Functional reconstruction of MLA draft models for speculative decoding", "family": "decodificacao_especulativa", "recommendation": "testar", "published": "2026-07-29", "as_of": "2026-10-18", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2607.27269", "independent_implementations": 0, "total_implementations": 0, "stars": 0, "citations": 0, "claims": [ { "claim_id": "claim-01", "statement": "The method replaces direct low-rank MHA/GQA-to-MLA conversion for draft models with a post-conversion calibration procedure that fine-tunes each converted MLA attention module to reproduce the original module's post-output-projection output on calibration hidden states, without verifier logits or supervision. Its cost is an additional per-module optimization stage after conversion, while the converted MLA cache layout and inference graph are preserved, so no extra serving-time memory or latency is introduced. It can fail when acceptance gains do not materialize (26 of 64 cells unchanged, one degraded), when calibration hidden states are unrepresentative of the production workload, or when the reader's stack uses MHA/GQA drafts directly and the MLA conversion step itself is not warranted.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2607.27269", "source_page": null, "source_excerpt": "" }, { "claim_id": "claim-02", "statement": "Across 64 matched task cells (four Llama/Qwen draft-target pairs, TransMLA and MHA2MLA converters, HF and vLLM backends, four 200-prompt tasks), functional reconstruction materially improves draft-token acceptance in 37 cells, leaves 26 unchanged within a 0.5-percentage-point tolerance, and decreases one; no aggregate speedup factor is reported.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2607.27269", "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": "Sx4.E11", "label": "(11)", "section": "Method", "role": "loss", "latex": "\\mathcal{L}_{\\mathrm{E2E}}^{(\\ell)}=\\mathbb{E}_{b}\\left[\\frac{\\left\\|M_{b}\\odot\\left(A^{\\ell}_{\\mathrm{MLA}}(H_{b}^{\\ell})-\\operatorname{sg}\\!\\left[A^{\\ell}_{\\mathrm{orig}}(H_{b}^{\\ell})\\right]\\right)\\right\\|_{F}^{2}}{D\\sum_{t}M_{b,t}}\\right].", "mathml": "ℒE2E(ℓ)=𝔼b​[‖Mb⊙(AMLAℓ​(Hbℓ)−sg⁡[Aorigℓ​(Hbℓ)])‖F2D​∑tMb,t].", "context": "For layer ℓ and calibration minibatch b, let Hbℓ∈ℝTb×D contain the cached layer inputs and let Mb∈{0,1}Tb mark valid tokens, broadcast over the hidden dimension. The implementation optimizes each layer independently with", "headline": "", "explanation": "", "objective": "", "symbols": {}, "evidence": "", "evidence_url": "", "source_url": "" }, { "anchor": "Sx3.E9", "label": "(9)", "section": "Preliminary: Why Converted MLA Is Not Draft-Preserving", "role": "metric", "latex": "\\alpha_{\\mathrm{sample}}(x) =\\sum_{y}\\min\\!\\left(p_{D}^{x}(y),p_{T}^{x}(y)\\right) \\\\ =1-\\mathrm{TV}\\!\\left(p_{D}^{x},p_{T}^{x}\\right).", "mathml": "αsample​(x)=∑ymin⁡(pDx​(y),pTx​(y))=1−TV⁡(pDx,pTx).", "context": "Speculative decoding makes this residual operationally visible. At a fixed context x, write pDx​(y)=pD​(y∣x) and similarly for pTx. Stochastic rejection-sampling verification has one-token expected acceptance (Leviathan et al. 2023; Chen et al. 2023)", "headline": "", "explanation": "", "objective": "", "symbols": {}, "evidence": "", "evidence_url": "", "source_url": "" }, { "anchor": "Sx3.E10", "label": "(10)", "section": "Preliminary: Why Converted MLA Is Not Draft-Preserving", "role": "metric", "latex": "\\alpha_{\\mathrm{greedy}}(x)=\\mathbf{1}\\!\\left[\\arg\\max_{y}p_{D}^{x}(y)=\\arg\\max_{y}p_{T}^{x}(y)\\right].", "mathml": "αgreedy(x)=[argmaxypDx(y)=argmaxypTx(y)].", "context": "Our experiments instead use greedy decoding, whose one-token acceptance event is", "headline": "", "explanation": "", "objective": "", "symbols": {}, "evidence": "", "evidence_url": "", "source_url": "" } ], "equations_status": "selected", "core_kind": "formula", "equations_fetched_at": "2026-09-15", "equations_source_url": "", "equations_source_sha256": "", "exposures_reviewed_at": "", "exposures_source_sha256": "" }