{ "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": "2607.25852", "title": "AngelSpec: Towards Real-World High Performance Inference with Speculative Decoding", "summary": "Replaces one universal speculative drafter with two specialized drafters, MTP for chat and block diffusion for code and mathematics, plus adaptive batched verification. It requires training both drafters for Hy3 target models and a custom server with a profiled verification-cost model. Benefits depend on trained drafters and control of the serving loop, which excludes third-party APIs and unsupported open models.", "rationale": "The reported gains require custom drafter training and serving control beyond the reader's single-GPU and third-party API setup.", "technique": "Hybrid MTP and block-diffusion speculative decoding", "family": "decodificacao_especulativa", "recommendation": "observar", "published": "2026-07-28", "as_of": "2026-10-18", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2607.25852", "independent_implementations": 3, "total_implementations": 3, "stars": 1160, "citations": 0, "claims": [ { "claim_id": "claim-01", "statement": "Replaces one universal speculative drafter with two specialized drafters, MTP for chat and block diffusion for code and mathematics, plus adaptive batched verification. It requires training both drafters for Hy3 target models and a custom server with a profiled verification-cost model. Benefits depend on trained drafters and control of the serving loop, which excludes third-party APIs and unsupported open models.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2607.25852", "source_page": null, "source_excerpt": "" }, { "claim_id": "claim-02", "statement": "Reports 1.98-2.40x speedup over autoregressive decoding and 10.5-11.8% higher throughput than DFlash on the Hy3 series.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2607.25852", "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": "S2.E9", "label": "(9)", "section": "2 Multi-Token Prediction", "role": "loss", "latex": "\\mathcal{L}_{\\mathrm{e2e}}=1-\\frac{1}{|\\mathcal{I}|D}\\sum_{i\\in\\mathcal{I}}\\sum_{m=0}^{D-1}\\prod_{k=0}^{m}\\alpha_{i,k},", "mathml": "ℒe2e=1−1|ℐ|​D​∑i∈ℐ∑m=0D−1∏k=0mαi,k,", "context": "After cold start, training switches to the end-to-end TV objective(Li et al., 2026) based on the expected accepted draft length, where αi,k=α⁡(pi,k,qi,k):", "headline": "", "explanation": "", "objective": "", "symbols": {}, "evidence": "", "evidence_url": "", "source_url": "" }, { "anchor": "S3.E12", "label": "(12)", "section": "3 DFly: Scalable and AR-Aware Parallel Drafting", "role": "proposed_method", "latex": "g_{t}^{(i)}=\\operatorname{RMSNorm}\\left(c_{t}+f_{t}^{(i)}\\right),", "mathml": "gt(i)=RMSNorm⁡(ct+ft(i)),", "context": "Motivated by their complementary roles, we formulate target conditioning as a shared-and-specialized decomposition. The FC branch ct establishes a common semantic basis by modeling cross-layer feature interactions, while ft(i) provides a depth-dependent refinement that preserves the target-layer preferences of draft layer i. The resulting conditioning representation is", "headline": "", "explanation": "", "objective": "", "symbols": {}, "evidence": "", "evidence_url": "", "source_url": "" }, { "anchor": "S4.E24", "label": "(24)", "section": "4 Real-World Deployment: D-cut", "role": "metric", "latex": "\\rho^{\\star}=\\arg\\max_{\\rho\\in\\mathcal{R}}\\frac{U(B,\\rho)}{C(B,\\rho)}.", "mathml": "ρ⋆=arg⁡maxρ∈ℛ⁡U⁡(B,ρ)C⁡(B,ρ).", "context": "At server startup, D-cut profiles the end-to-end speculative-step latency C⁡(B,ρ) for each supported (B,ρ) pair, including drafting, selection, packing, and target verification. At runtime, it chooses", "headline": "", "explanation": "", "objective": "", "symbols": {}, "evidence": "", "evidence_url": "", "source_url": "" } ], "equations_status": "selected", "core_kind": "formula", "equations_fetched_at": "2026-09-14", "equations_source_url": "", "equations_source_sha256": "", "exposures_reviewed_at": "", "exposures_source_sha256": "" }