PaperlayerView exact revision on arXiv

Source-linked AI summary

Phase-Scoped Query Layouts for Reference-Constrained Speculative Decoding

Chengzhan Li

arXiv:2610.03228v4cs.PF

TL;DR

Speculative verification changes a quantized target model’s execution geometry and can change its greedy output, creating a challenge when the deployed output is the required reference. The paper introduces phase-scoped query-to-batch attention and evaluates its latency, numerical correspondence, and conditions for deployment.

  • Problem

    Speculative verification evaluates multiple query positions together, and finite-precision execution can change a quantized deployment’s greedy output.

    Source: §1 · p. 1

  • Method

    Phase-scoped QB pairs query and mask views to represent queries as attention batch instances while keeping surrounding target computation batched.

    Source: §1 · p. 1, §3.1 · p. 2

  • Results

    3.06%, 6.17% and 2.83% lower speculative request latency for QB than PQ are reported for Qwen3, SmolLM2 and Llama, respectively.

    Source: §1 · p. 1

  • Takeaways & Limitations

    The design keeps local numerical correspondence, stateful reference fidelity and complete-request cost as distinct, testable deployment obligations.

    Source: §10 · p. 14

    A 737-token Llama input shows that PQ and QB can agree with each other yet differ from the deployed reference.

    Source: §1 · p. 1, §2.3 · p. 2

Abstract

from arXiv · show

Speculative verification changes the execution geometry of a quantized target model and can change its greedy output. We present phase-scoped query-to-batch (QB) attention, a graph transformation that pairs query and mask views, shares key/value storage, and retains batched projections and feed-forward computation. Phase-aware dispatch and layout-sensitive graph reuse bind the transformation to the native speculative transaction. In a six-session CUDA study, QB reduces request latency relative to explicit per-query attention (PQ) by 3.06%, 6.17% and 2.83% for Qwen3, SmolLM2 and Llama; a 72-context replay finds bitwise correspondence at QK, softmax and AV. A matched seven-arm study shows that local rather than global repair changes draft profitability for two families. A prospective transfer study then tests 36 model/input cells drawn from 21 previously unused tasks. SmolLM2 preserves the reference on its complete panel and obtains a 6.99% layout reduction; Qwen3's 3.22% estimate is noise-limited. On a 737-token Llama input, PQ and QB agree with each other but differ from the deployed reference, separating local numerical correspondence from end-to-end preservation. Vulkan state interventions and a scoped AV-preserving port identify producer-state and backend obligations. Complete-service studies further distinguish representation savings from the net value of drafting. The results establish a local execution method with measured cost benefits and distinguish its numerical, stateful and complete-cost conditions for deployment.

1. Introduction

The paper asks how to preserve a deployed quantized model’s greedy output during speculative verification without giving up all batching. It introduces phase-scoped query-to-batch attention and specifies the numerical, state, and graph-reuse conditions for applying it.

1. Introduction

  • Speculative verification evaluates several query positions together, and finite-precision execution can therefore change a quantized deployment’s greedy output.The study treats that deployment’s output as the required reference while asking how much parallel work can remain.Source: §1 · p. 1

3.1. Separating attention from target-wide batching

  • QB pairs query and mask views to represent queries as separate attention batch instances, while sharing K/V and retaining batched surrounding computation.The native speculative acceptance and state-reconciliation protocol remains unchanged.Source: §1 · p. 1, §3.1 · p. 2, Figure 1 · p. 3

1. Introduction

  • 3.06%, 6.17% and 2.83% lower speculative request latency for QB than PQ are reported for Qwen3, SmolLM2 and Llama, respectively.The paper distinguishes this incremental graph-lowering comparison from the broader benefit of local repair versus global serialization.Source: §1 · p. 1

2.1. Output qualification

  • Output qualification requires complete candidate endpoints to match a stable reference on the registered panel, configuration and output window.The paper distinguishes endpoint agreement from exact-arithmetic equivalence and bitwise operator correspondence.Source: §2.1 · p. 2

2.2. The stateful transaction

  • A speculative transaction tracks target and draft state, evaluates pending and proposed tokens, commits accepted output, and reconciles rejected positions before the next cycle.A visible token may still be pending evaluation into KV and must be consumed exactly once.Source: §2.2 · p. 2

2.3. Local repair and state composition

  • Local PQ/QB correspondence on one input state does not establish agreement with the reference when upstream execution forms a different state.Full stateful composition requires a relation that holds initially, preserves committed output, and holds again after reconciliation.Source: §2.3 · p. 2

3.1. Separating attention from target-wide batching

  • QB’s query view preserves the original query values and stride while changing their logical shape from (d, q, hQ, 1) to (d, 1, hQ, q).The corresponding mask view retains query-dependent values, and K/V retain their native mapping and shared single-stream ownership.Source: §3.1 · p. 2, §3.2 · p. 2, A Attention organization versus… · p. 3

3.3. Admission conditions are part of the method

  • Admission guards constrain the supported graph, while phase selection and active layout must be reflected in graph parameters and reuse checks.The prior finite mask-family test selected mask 3 as inclusion-minimal, but its separate cost comparison did not pass the practical speed threshold.Source: §3.3 · p. 3, §3.4 · p. 4, §3.5 · p. 4

4. Experimental design

The experiments separate output preservation, local numerical correspondence, repair scope, and complete-request cost. They use distinct panels and controls for those questions rather than treating repeated measurements as independent tasks.

4.1. Systems, panels and distinct evidence roles

  • The principal campaigns use three Q8_0 model pairs on CUDA, with a separate Vulkan setup for state and portability experiments rather than controlled GPU-speed comparisons.The CUDA reference uses physical UB1; repaired execution uses target caller UB4, draft UB1, k = 3, F16 KV and non-flash attention.Source: §4.1 · p. 4

4.2. Matched alternatives and complete-request timing

  • The complete nonstreaming resident-process HTTP timer includes encoding and submission, generation, response parsing, and endpoint validation; model loading and registered warmups are outside it.Diagnostic times are separate and are not subtracted from request measurements.Source: §4.2 · p. 4
  • E2 uses fresh processes, short counterbalanced PQ/QB blocks, and six sessions spanning 24.45 hours; its 95% interval bootstraps six session-level log contrasts.The interval describes a fixed panel on one host, not six independent dates; the PQ/QB practical ratio threshold is 1.02.Source: §4.2 · p. 5

4.3. Numerical and state controls

  • The 72-context replay tests three layer positions, prefill and continuation, and query widths one through four after observation, native-MQ and repeatability controls.Contexts were replayed twice per arm; the numerical claim is withheld if a corresponding control fails.Source: §4.3 · p. 5, Figure 2 · p. 6
  • Table 3 presents a larger frozen-panel confirmation, reporting matching complete endpoints out of 240 per arm/family; M0 denotes unconstrained UB4, and Rprod an engineering-selected ordinary recipe.Source: Table 3 · p. 6

5. Numerical correspondence and repair cost

The section establishes that repair organization affects both reference preservation and drafting economics, while QB retains local attention correspondence and lowers request cost relative to PQ. Results distinguish the benefit of local repair from the smaller incremental benefit of QB’s representation.

5.1. Complete-output qualification across model families

  • Repaired target-only M3 and speculative C3 match all 240 required endpoints per family, while unconstrained M0 fails on 5, 14 and 9 distinct tasks for Qwen3, SmolLM2 and Llama.Ordinary Rprod comparisons for SmolLM2 and Llama are not reported as reference-preserving deployment gains because that recipe fails for those families.Source: §5.1 · p. 6

5.2. Same-input correspondence across layers and phases

  • Across 72 contexts, PQ and QB are bitwise equal at QK, softmax and AV, with zero maximum error and ULP distance.MQ differs first at QK in 54 multi-query contexts; sampled visible contexts reach at most 109, 130 and 127 keys by family.Source: §5.2 · p. 6

5.3. QB’s independent implementation benefit

  • QB lowers speculative request latency than PQ by 3.06%, 6.17% and 2.83% for Qwen3, SmolLM2 and Llama, respectively.All six-session speculative intervals exceed the 1.02 practical threshold; all 36 family/path/session contrasts favor QB, though 9 of 72 family/path/task aggregates regress.Source: §5.3 · p. 6

5.4. Repair scope changes draft profitability

  • Speculative GS/QB ratios are 2.144, 1.463 and 2.176, while PQ/QB ratios are 1.044, 1.068 and 1.022 for Qwen3, SmolLM2 and Llama.The first comparison reflects local versus global repair; the second isolates the representation choice within local repair, rather than successive software-version gains.Source: §5.4 · p. 6
  • Adding drafting to QB is profitable for Qwen3 and Llama but not SmolLM2: MQB/CQB is 1.144, 0.634 and 1.117, respectively.Under GS, drafting is unprofitable for Qwen3 and Llama; both cross parity under local repairs, while SmolLM2 remains below parity.Source: §5.4 · p. 7, Figure 4 · p. 7
  • Target-only QB preserves the specified serial reference at lower cost, with R/MQB ratios of 1.334051, 1.369890 and 1.426161 across the three families.Thus SmolLM2 benefits from reference-preserving batching even though drafting is unprofitable.Source: §5.4 · p. 7

6. Prospective reference and workload transfer

A prospective study tests reference qualification and layout cost on held-out inputs, separating local PQ/QB correspondence from fidelity to the deployed reference. SmolLM2 clears the registered layout screen, Qwen3’s estimate is noise-limited, and Llama’s reference mismatch leaves its timing unqualified.

6.1. A frozen test beyond the development inputs

  • The sealed prospective panel contains 36 family/input cells from 21 tasks, with four natural prompts per family in each of three token-length strata.The short stratum uses GSM8K mathematics and the two longer strata use CodeContests descriptions, so workload and prompt length change together.Source: §6.1 · p. 8

6.2. Qualification separates correspondence from reference fidelity

  • Qwen3 and SmolLM2 qualify on all twelve inputs, but Llama departs from the deployed reference on a 737-token prompt at output index 81.PQ and QB agree in complete decision-vector bytes, canonical causal trace and output; the remaining Llama inputs and timing slots are unstarted under the stop rule.Source: §6.2 · p. 8, Figure 5 · p. 8
  • The Llama counterexample shows that agreement between local implementations does not establish fidelity to the deployed reference.The cause of their shared departure is not isolated, and the observed prompt length is not established as a universal threshold.Source: §6.2 · p. 8

6.3. Held-out layout benefit and complete-request cost

  • SmolLM2’s QB-to-dual-PQ ratio is 0.930056, a 6.99% reduction that clears the practical and A/A thresholds; Qwen3’s 0.967781 ratio is a 3.22% reduction below its 4.78% A/A threshold.Figure 6B shows the measured input/session ratios and registered cost thresholds; eight of Qwen3’s 48 blocks favor PQ.Source: §6.3 · p. 8, Figure 6 · p. 9
  • Relative to serial R, Qwen3’s QB and dual-PQ ratios are 0.560577 and 0.579239, while SmolLM2’s are 1.007007 and 1.082739.These comparisons change both target organization and drafting, unlike matched contrasts that isolate the draft increment under one repair policy.Source: §6.3 · p. 8, Figure 6 · p. 9

7. State formation and transfer boundaries

State interventions show that local correspondence depends on the state produced upstream, while backend admission and AV evaluation impose additional constraints on transfer. Controlled substitutions and a scoped port identify both supported behavior and limits.

7.1. The input to a local repair matters

  • For the two exposed divergent inputs (161 and 133 prompt tokens), B reproduces C’s complete endpoint, initial FP32 logits, and sequence-KV bytes, while A differs at generated indices 47 and 81.A partition sweep produces several equivalence classes, including a third trajectory for the 133-token input split as 132 + 1. Smaller chunks are not monotonically closer to the reference.Source: §7.1 · p. 9
  • First-layer Q/K/V projections already differ across selected prefill organizations, while eight scale/MUL replay calls are excluded because captured operands were overwritten.The admitted fixed-operand replay qualifies 32 of 40 calls; a driver-body preregistration gap also limits the partition sweep.Source: §7.1 · p. 9
  • In a validated receiver, suffixes follow the KV source, while swapping initial logits is a negative control because all initial vectors choose token 1271.Stable controls retain their endpoint despite unequal KV, showing that local consumer correspondence on one state does not establish correspondence on every upstream-produced state.Source: §7.1 · p. 9

7.2. Endpoint agreement and state substitution

  • Across 160 state substitutions, layer-0 K alone or V alone—1/64 of K/V storage—reproduces a donor endpoint in some coordinates, but enlarging a replacement can remove that effect.The tested coordinates include 30 donor endpoints, 54 third trajectories, and 76 recipient endpoints despite changed bytes.Source: §7.2 · p. 10
  • Stable inputs can have equal complete endpoints despite unequal KV, and substitutions produce third trajectories on 16/20 coordinates for /1011 but none on /1249.Only 44 of 112 planned repeat-level contrasts have all required historical vectors; the interventions establish finite counterexamples, not natural reachability or a globally minimal repair.Source: §7.2 · p. 10
  • For /1248 B→A at output 47, the additive joint score prediction is +0.081377 but the actual joint score is negative, with IKV = −0.135864.The four-cell argmax pattern is 220/220/832/220, exhibiting net score nonadditivity through subsequent computation.Source: §7.2 · p. 10

7.3. A legal view can fail backend admission

  • The natural-execution pilot has incomplete diagnostics and changed two QB outputs; separately, q = 2 violates a necessary condition of the frozen Vulkan matrix-multiplication backend.The pilot supplies endpoint counterexamples but no qualified speedup; actual operator placement and fallback transfers remain unobserved, so it cannot qualify an all-Vulkan numerical comparison.Source: §7.3 · p. 10

7.4. Preserving the AV path in a scoped GQA port

  • The scoped packed AV route differs from PQ in 7,160 of 8,192 values at key512/q = 2 and fails exact reference, while an AV-preserving route matches all six tested coordinates byte-for-byte through final output.The AV-preserving route retains per-output evaluation and uses zero V batch stride; its real-request qualification is separate.Source: §7.4 · p. 10

8. Complete transaction costs

Complete-transaction measurements separate cheaper mapped execution from practical end-to-end benefit: QB layout gains are unresolved in the service panels, while drafting can still reduce cost relative to target-only execution on one panel. Representation and path changes must qualify against the reference and be charged across the transaction.

8.2. Path availability and reference preservation

  • Padding logical length 1023 to physical extent 1024 reaches the three-kernel path on the observed first call, but all four hosts fail the original reference.Padding to 1025 passes all four hosts, retains eight kernels, and costs about 1.5% more than either native QB control; the 8/3/8-kernel observation is not a qualified steady-state comparison.Source: §8.2 · p. 11, Figure 8 · p. 12

8.3. Charging the representation transaction

  • Compact-QH/QB_A transaction ratios are 1.002216 at q = 2 and 1.001598 at q = 4; neither candidate passes the practical screen against both controls.The charged transaction includes producing and consuming the representation; the B570 adjusted interval [0.979908, 1.004691] likewise establishes no QH advantage.Source: §8.3 · p. 11, B Complete Q-representation transactions · p. 12, Figure 8 · p. 12

8.4.1. Service and reference contracts

  • The registered complete-service window charges context creation, graph reservation, prefill, generation, sampling, state handling, synchronization, and context destruction; input formatting and policies differ across platforms.Those platform differences mean the comparison does not isolate the GPU, and the service ratios do not replace the separate resident-request estimates.Source: §8.4.1 · p. 11

8.4.2. Vulkan request qualification and layout value

  • The QB/dual-PQ service ratio is 0.995144, a 0.486% point reduction below the 2% practical margin, while QB/target-only service is 2.619763 on the Vulkan panel.All required endpoints and complete cycle traces agree, but the tested speculative policy costs more than target-only service.Source: §8.4.2 · p. 11
  • A draft CPU-thread change from 8 to 1 preserves all 346 original vectors but yields a 1.003994 service ratio, establishing neither a speedup nor the physical cause of draft cost.The target-verification host interval falls about 7.33%, while draft prefill/decode costs 12.366798 seconds per QB request.Source: §8.4.2 · p. 11

A Fixed logical length 1023: RTX

B Complete Q-representation transactions

  • The B570 complete Q-representation transaction shows no established practical QH benefit; its descriptive adjusted interval is [0.979908, 1.004691].The RTX and B570 rows use different geometries and execution modes, so their ratios are not a device comparison.Source: B Complete Q-representation transactions · p. 12, Figure 8 · p. 12

8.4.3. RTX: useful speculation, unresolved layout increment

  • On RTX, the layout ratio is 0.989284, a 1.072% point reduction that remains unresolved; QB/target-only service is 0.582325, while dual-PQ/target-only is already 0.588633.The 41.8% lower QB service time is not attributable to QB, and the declared practical screen finds no established layout improvement.Source: §8.4.3 · p. 12

9. Design implications and relation to prior work

Deployment should be qualified against producer-state and backend requirements, complete requests, and complete costs; local query-to-batch gains do not alone establish deployment fidelity or service value. The evidence is limited to specified panels and execution settings, distinct from related systems’ guarantees and comparisons.

9.1. The implementation boundary that transfers

  • Qualification should establish producer-state and backend contracts, test complete requests, and compare complete costs under the frozen deployment.The address argument and AV-preserving Vulkan port provide distinct implementation checks at the graph/backend interface.Source: §9.1 · p. 12
  • Complete transaction costs can absorb QB’s incremental representation benefit or leave drafting unprofitable, so local savings do not establish net request value.The negative submission control also separates host API granularity from physical query width and measured service cost.Source: §9.1 · p. 13

9.2. Related work

  • QB changes attention representation without adding a numerical-consistency verifier; external systems provide design context, not measured speed comparisons.The paper distinguishes its finite qualification against an existing deployment from engine/kernel guarantees under stated assumptions [27].Source: §9.2 · p. 13

9.3. Scope of the evidence

  • CUDA gains are qualified only for one logical sequence, non-flash attention, occupied q ≤4, Q8_0 weights, F16 KV, and supported model paths.Larger contexts, concurrency, other precisions, fused attention, and comparison with an optimized fused-attention stack require separate qualification.Source: §9.3 · p. 13
  • Service panels use four inputs, one device/date, and do not establish equivalence; the prospective panel is finite and leaves three Llama cells unobserved.Neither panel establishes universal determinism or a production serving ceiling; production throughput, energy, and hardware efficiency also remain outside measurement.Source: §9.3 · pp. 13–14

10. Conclusion

The conclusion presents phase-scoped QB as a local way to separate attention queries while retaining batched surrounding work, with measured correspondence and lower cost than PQ. Transfer results and controls make producer state, backend behavior, reference fidelity, and total cost separate deployment obligations.

10. Conclusion

  • QB pairs query and mask views, shares K/V semantics, restores output order, and reuses graphs according to layout while retaining batched surrounding computation.Controlled CUDA experiments establish local correspondence and lower request cost than PQ; matched repair comparisons show repair policy and draft decisions should be considered together.Source: §10 · p. 14
  • SmolLM2 retains a measurable layout gain, Qwen3’s increment is noise-limited, and Llama shows local PQ/QB agreement can coexist with failure to preserve the deployed reference.Vulkan interventions and complete-service controls place producer state, backend execution, and request-boundary cost within qualification.Source: §10 · p. 14

S1.2. Complete phase-selection experiment

  • Mask3 was inclusion-minimal only within the tested phase-selection family, while the separate all-phase comparison did not establish an efficiency improvement.The all-phase comparison measured 1.003569 generation speed, below its registered 1.03 threshold, with an interval including parity.Source: S1.2. Complete phase-selection… · p. 16

S1.3. Width and runtime boundaries

  • Widening k from 3 to 7 preserved endpoints but reduced generation/request ratios to 0.829442/0.889868; the observed maximum reference length was 261 tokens.Requesting cap512 did not establish a 512-token complete trajectory, and the wider geometry had failures on MBPP229.Source: S1.3. Width and runtime boundaries · p. 16

S2.1. Larger confirmation

  • The larger confirmation included 3,600 complete requests, with M3 and C3 matching 240/240 endpoints per family; ordinary-production qualification failed on 140 SmolLM2 and 80 Llama requests.Regressed tasks for M3/C3 numbered 0/24, 24/24, and 4/24 across families; failed comparisons remain unqualified.Source: S2.1. Larger confirmation · pp. 16–17, Table S2 · p. 17

S2.2. Seven-arm implementation comparison

The seven-arm comparison finds that local repair changes draft profitability relative to global serialization, while QB reduces the remaining representation cost relative to PQ. Separate six-session timing and operator replay studies test the PQ/QB increment and local correspondence.

S2.2. Seven-arm implementation comparison

  • Qwen3 and Llama move from unprofitable drafting under GS to profitable drafting under both local repairs, while PQ already crosses that boundary and QB reduces remaining representation cost.SmolLM2 favors QB on all tasks in both local comparisons, and every GS/QB task favors QB; drafting regresses on all 12 SmolLM2 tasks.Source: S2.2. Seven-arm implementation… · p. 17, Table S5 · p. 17

S3.1. E2 design and complete estimates

  • Six-session E2 reports task-level PQ/QB ratios across the retained 12-task panel, with 9 of 72 task aggregates regressing.The intervals describe fixed-panel, single-host variation, not date-level population variation or a multiple-comparison guarantee.Source: S3.1. E2 design and complete estimates · pp. 17–18
  • E2’s recorded acceptance, rollback, and logical-workload counters agree across all 216 speculative task/session pairs, but complete control-trace identity was not tested.Model callers and physical micrographs remain distinct from hardware launches.Source: S3.1. E2 design and complete estimates · p. 18

S3.2. E1 coverage and controls

  • PQ and QB are bitwise equal at all three operator stages in all 72 contexts; MQ/PQ and MQ/QB equality occurs only in the 18 q = 1 controls.The replay compares operator arrays, not next-token logits.Source: S3.2. E1 coverage and controls · p. 18

S3.3. Restoration and interpretation

  • Restoration reconstructed 11 analysis outputs byte-for-byte on the same host, but did not externally replicate the study or replay the original tensor archives.The manuscript checker reconstructs ratios and intervals from the submitted session table.Source: S3.3. Restoration and interpretation · p. 18

S4. Vulkan partition and retained-state experiments

The Vulkan studies compare partitioned and whole-prefill states, reconstruct projection and boundary behavior, and test retained-state interventions. Their results are conditional on the tested inputs, receiver, and execution path; the cost evidence does not supply a completed component-time model.

S4.1. Inputs and partition controls

  • Across the original partition inputs, B/C full endpoints and initial FP32 logits and KV payloads agree; the schedule table records first output differences from A without implying numerical monotonicity.The selected partition classes distinguish whole prefill from a 128-split class, not a universal 128-token threshold.Source: Table S9 · p. 19, Table S10 · p. 19, Figure S2 · p. 20

S4.2. Projection and boundary reconstruction

  • Sixteen own-path shams pass before 40 crosses, supporting conditional suffix sensitivity but not a general opaque-state clone.The boundary0 cross schematic reports that initial logits select 1271 and subsequent suffixes follow the KV source under the same reconstructed receiver.Source: S4.2. Projection and boundary… · p. 20, Figure S3 · p. 20
  • Thirty-two of 40 projection calls pass native same-arm correspondence; eight scale/MUL calls fail because captured inputs alias buffers already changed in place.Admitted differing projection replays reproduce geometry sensitivity, while RMS and the first task’s Q controls agree bitwise.Source: S4.2. Projection and boundary… · pp. 19–20, Table S11 · p. 20

S4.3. Separate physical-work and cost boundary

  • The post-hoc production-path pairing reports Rprod/C = 0.719264 and Mmatch/C = 0.727927, with every task regressing; it does not repair the failed original stock-relative qualification.No completed component-time model is admitted.Source: S4.3. Separate physical-work and cost… · p. 20

S5.1. Selectors and complete endpoint matrix

  • The endpoint matrix preserves recipient logits and pending-token consumption while varying selected donor KV components; all 76 D requests change actual bytes.K-only or V-only layer0 selects 1/64 of K/V value storage, excluding metadata, and selectors follow actual K and V storage strides.Source: S5.1. Selectors and complete endpoint… · p. 20

S5.2. Magnitude and common-prefix selection

The appendix tests whether perturbation magnitude explains endpoint divergence, audits direction symmetry, and reports the scope of available captures. It also records a stopped exploratory service pilot and an independent representation audit.

S5.2. Magnitude and common-prefix selection

  • F16 perturbation magnitude does not perfectly separate endpoint labels; the first-divergence frontier is output-specific, and the archive supports only 44 of 112 planned repeat-level contrasts.For /831 recipient A, the L0 and POSITION frontiers are outputs 40 and 81; later vectors cannot supply an earlier missing comparison.Source: S5.2. Magnitude and common-prefix… · pp. 20–21

S5.3. Direction symmetry and independent conditions

  • Reversing donor and recipient permutes the POSITION quartet, while L0 retains distinct surrounding recipient states; the direction audit covers 12 available pairs across four underlying inputs.For /1248 at output 47, both directions give −0.09655952453613281 in pair order (220, 832), despite opposite original signs.Source: S5.3. Direction symmetry and… · p. 21

S5.4. Acquisition failure and data scope

  • The planned EMPTY B→A acquisition imported the wrong recipient, so its 127 decodes are not qualified samples and 167 slots remained unstarted.The later CPU revision repairs assembly but adds no native data or missing frontier comparisons.Source: S5.4. Acquisition failure and data scope · p. 21

S6.1. Design and complete exploratory endpoints

  • The natural pilot stopped after 49 attempts, leaving P2/P3 unrun and qualified speedup null; its monitoring declaration was unmet and candidate observer pairs were incomplete.The q1 branch bypassed multi-query QB, so draft-free QB-labelled agreement does not qualify lifted QB; the missing multirow witness neither proves nor disproves that such execution is possible.Source: S6.1. Design and complete exploratory… · pp. 21–22, Table S15 · p. 22

S6.2. Independent representation audit

  • The CPU representation audit detected all four injected faults and had no strict assertion, crash, or nonfinite failure, but its dyadic inputs do not establish generic F32 reduction invariance.General maximum absolute error was 5.3608031×10−9 at softmax and 5.5132505×10−5 at AV/final output; endpoint relations do not establish common operands, KV, or acceptance traces.Source: S6.2. Independent representation audit · p. 22

S6.3. Version-bound support semantics

The section identifies backend and measurement conditions that constrain representation-level and service-cost claims. It reports software and graph interventions, fixed-input cost windows, request-observer checks, and two distinct service panels.

S6.3. Version-bound support semantics

  • The frozen Vulkan support predicate rejects the reviewed QK and AV shapes; the source study adds no GPU calls, observed fallback, or retroactive pilot qualification.Matching fourth-axis extents alone would not certify the full support predicate.Source: S6.3. Version-bound support semantics · p. 22

S7.1. Input lifetime and software controls

  • Input-lifetime interventions localized the RTX overwrite of K/V storage, while the QB_A/QH ratio was 0.900826 across the core q = 2, 4 panel and 19 of 24 geometries descriptively favored QB.The selector predicted all 12 held-out synthetic mean directions but cost 1.010298 times fixed QB after lookup cost.Source: S7.1. Input lifetime and software… · p. 22

S7.2. Fixed-logical-input padding and cost windows

S7.3. Request relevance and observer integrity

  • The first B570 request bridge was a distinct q1/key256/GQA32:8 observation; serial logging added about 12% median decode overhead and failed its observer screen.The repaired same-DLL clean/profile/clean medians were 27.6675/29.6058/27.6794 ms, so the observer still added about 7%.Source: S7.3. Request relevance and observer… · p. 23

S8.1. Panel identities and qualification

  • The RTX and B570 service panels used separate input sets and contracts, and their service ratios cannot be treated as paired GPU comparisons.RTX timing retained complete speculative traces, while the B570 AV repair qualified a separate four-input panel before timing.Source: S8.1. Panel identities and qualification · p. 23

S8.2. Every input and paired block

  • RTX QB/R blocks all fell below one, whereas B570 blocks all exceeded one; aggregate ratios were 0.582325 and 2.619763, respectively.RTX dual-PQ/R was already 0.588633, so its lower QB/R service time does not isolate the layout benefit; neither platform passed the practical layout screen.Source: S8.2. Every input and paired block · pp. 23–24, Table S17 · p. 24
  • Changing B570 draft context parameters from 8 to 1 left all 346 vectors and endpoint/state traces unchanged, but DRAFT1/dual-baseline service was 1.003994 and did not support practical improvement.The 64 measurements across four timing processes had an A/A criterion of 0.793%.Source: S8.2. Every input and paired block · p. 25

S8.4. Evidence and execution accounting

The evidence accounting separates host-reported archives from the manuscript reconstruction and records the scope of a prospective transfer study. That study reports preservation outcomes for qualified inputs, a full-reference counterexample, and unqualified comparator failures.

S8.4. Evidence and execution accounting

  • The RTX archive contains 3,110 members and 2,900 vectors, but the manuscript package does not include that full native archive; local restoration, compact-record reconstruction, and independent raw access are distinct evidence levels.The originating host reports reopening and hashing every member.Source: S8.4. Evidence and execution accounting · p. 25

S9.1. Frozen identities and the natural-input amendment

  • The amended prospective study selected 36 model/input cells from 21 tasks and executed 33 cells across 20 tasks: 32 preserved the deployed reference, one did not, and three Llama cells remained unstarted.These coverage counts are not independent trials or an estimated population success rate; sequence length is confounded with domain in the prompt bins.Source: S9.1. Frozen identities and the… · p. 25

S9.2. Observation and qualification semantics

  • On the 737-token Llama input, PQ and QB agreed in decision vectors, causal trace, and output, but both first differed from the serial reference at output 81.This establishes a full-reference boundary without isolating its numerical cause; the eight earlier qualified inputs do not form a replacement Llama cost panel.Source: S9.2. Observation and qualification… · p. 26
  • The NATIVE_FAST route failed calibration or held-out reference checks for all three families, so its timing remained unrun and it supplies counterexamples rather than a qualified cost victory.The failure indices were 8 and 22 for SmolLM2 and Llama calibration, and 68 for Qwen3 held-out evaluation.Source: S9.2. Observation and qualification… · p. 26

S9.3. Complete held-out timing and estimation

Complete held-out timing found a practical layout reduction for SmolLM2 but not Qwen3. A separate B570 submission-policy closeout found no practical benefit, while the evidence package distinguishes reproducible reported results from full raw-data access.

S9.3. Complete held-out timing and estimation

  • SmolLM2 passed the practical layout screen, while Qwen3 did not; their session ratios were 0.928001–0.931868 and 0.960604–0.975669, respectively.The screen required reduction beyond both A/A imbalance and 2%, with every session favoring QB; no Llama cost was reported.Source: S9.3. Complete held-out timing and… · p. 27

S9.4. B570 one-factor submission-policy closeout

  • B570 SUBMIT/dual-BASE was 0.999161301690, a 0.0839% reduction that missed the 2% floor and reversed in one session; the study did not establish practical benefit or equivalence.The service ratios SUBMIT/R and dual-BASE/R were 1.203677993457 and 1.204688363552.Source: S9.4. B570 one-factor… · p. 27, Table S20 · p. 27
  • The B570 closeout retained an accounting deviation: auxiliary descendants were not completely counted or capped, so full protocol resource compliance was not certified.The result was accepted as a limited negative closeout.Source: S9.4. B570 one-factor… · p. 27

S10. Reproduction and evidence access

  • The public package reconstructs reported ratios and plotted values but does not provide complete raw archives, native inference, or model weights; accessible evidence routes remain necessary for external review.Manuscript-level reconstruction and originating-host raw-data audits are separate evidence levels.Source: S9.4. B570 one-factor… · p. 28, S10. Reproduction and evidence access · p. 28
Loading 2610.03228v4…