Repository navigation
build: pin ns-cmsis-nn v7.40.0 #426
Description
ns-cmsis-nn v7.40.0 is being cut from main 07fb2eee plus a version bump. helia-rt pins v7.39.3 (94bf3b93).
Plan
- Move the pin in
helia.inc, the download fallback, the canary default,smoke_cmake.yml'sNS_CMSIS_NN_REF, thekernels/heliaREADME and the Zephyr guide. - Prepare against main
07fb2eeenow and swap to the tag SHA when it is cut.NS_CMSIS_NN_REF(the release pack) moves with the tag. - Pin bump only. New heliaCORE entry points are not wired here.
- Acceptance:
- the helia suites pass on all legs, with the counts explained;
- the CMSIS-Pack floor is checked;
- an HPX board A/B (helia backend, GCC and ATfE) runs against main under the no-regression rule;
- the PR publishes at the tag SHA with one CI cycle.
- Worktree:
teams/helia/workspaces/rt/worktrees/helia-rt-pin-740, branchrt/pin-ns-cmsis-nn-7.40.0, team helia, lane RT.
Metadata
Metadata
Assignees
Labels
No labels
Activity
apage224 commentedon Oct 5, 2026
Checkpoint (handoff)
cf2ac068on branchrt/pin-ns-cmsis-nn-7.40.0, local and not pushed, on maine1f49019. Two commits:dc2c4e1apins v7.40.0 (b8102da7) at every pin site, including smokeNS_CMSIS_NN_REF; the CMSIS-Pack floor stays 7.39.2;cf2ac068makes the heliaconvolution_status_testcount only the kernel's own weight-sum size queries, since v7.40.0's MVE precompute queries the size itself (ns-cmsis-nn #687).e1f49019vs900e41ce, the code-identical pre-tag head), which is queued;apage224 commentedon Oct 5, 2026
Resume checkpoint
Head: cf2ac068525e477e71d4afd96f5007ce988811dc, local branch rt/pin-ns-cmsis-nn-7.40.0.
Scope: fold the heliaCORE comment spelling nit, finish fresh independent reviews, and reconcile HPX measured-head equivalence. Acceptance remains the existing seven-leg correctness evidence plus required board A/B under the no-regression rule. Publication READY and one CI cycle only after the A/B passes.
Worktree: /home/adamp/Ambiq/aitg/orchestration/teams/helia/workspaces/rt/worktrees/helia-rt-pin-740; team helia, lane RT.
apage224 commentedon Oct 5, 2026
Board validation resumed for requested head
d0e07edf2f5b24343a1f05cddc2290dfe34ec472, retaining B image source900e41ce9f29cf1de682ca9d47e666b8546081f4vs Ae1f490191f23c83e45a799c2461b25af4149971f. Source/compile-condition and archive/image hash reconciliation:artifacts/hpx/rt-426-ab/RECONCILIATION.mdandIMAGE-PROVENANCE.json. Scope: remaining Apollo510B GCC/ATfE 39-model matrix and Apollo4 GCC 15 integer models; reuse completed captures. Report output deltas, model changes over 1%, layer changes over 3%, and limitations. Physical claim per run; AOT M4 queue first. Team/lane helia/HPX. No source edits, publication or CI.apage224 commentedon Oct 5, 2026
Checkpoint: local validation complete; board gate pending.
Head: cf0490b1102ec145f5f0fbb66c511a863c1543de on rt/pin-ns-cmsis-nn-7.40.0, local and unpublished. The heliaCORE spelling nit and a released Zephyr example mismatch are fixed; retained pre-tag refs and worktrees remain intact.
Done: three fresh independent correctness reviews passed. Final light found that the Zephyr example claimed to match released v2.0.0 while specifying a newer dependency; the docs-only fix restores its actual v7.39.2 pin and passed affected-delta review. Active build/CI pins remain v7.40.0. A separate final light passed at cf0490b1. Targeted M55 GCC SPEED FVP evidence: previous instrumentation fails 12/24 cases; final instrumentation passes 24/24; inserting an extra real size query into conv/depthwise again fails 12 cases at the exact count assertion. Six local workflow gates and whitespace/shell/YAML/format checks pass. Seven broader suite logs remain inherited from RT 900e41ce / CORE 07fb2eee, with algorithm-source equivalence documented; they are not final-tag artifact or CI runs.
Remaining: HPX must reconcile retained firmware with this exact head and complete required board A/B. Its preceding AOT work is still in progress. No performance/no-regression claim yet. Publication READY + one ordinary CI cycle remains conditional on that pass; no push, PR, CI, merge or release has occurred.
Next: incorporate HPX results, recheck remote main/final head, publish under the existing conditional authorization, observe the single CI cycle, and return the merge card. Local card/evidence: artifacts/rt/pin-ns-cmsis-nn-740/ in orchestration.
apage224 commentedon Oct 6, 2026
Board A/B complete: 186/186 retained images, 93/93 full-output pairs byte-identical and finite. Apollo510B: 39 models with GCC14.3.1 and ATfE22.1; Apollo4: 15 integer models with GCC. LP96MHz, effective -Ofast, identical placement per A/B.
Whole-model ranges: M55 GCC -1.73%..+0.47%; M55 ATfE -0.42%..+0.23%; M4 GCC -10.88%..0%. No positive model flag exceeds1%. However, 49 positive layer flags exceed3% (25 with either median≥1000 cycles,24 smaller). Examples: M55 GCC IC int8 FC1100→1298 (+18%, spreads99/187); VWW int8 conv145115→159855 (+10.16%, variable); M4 arr int8 conv1704→1867 (+9.57%, zero sample spread). These are initial flags without independent repeats, requiring explicit performance disposition; this is not a blanket no-regression result.
Linked text ranges: M55 GCC -1440..0 B; M55 ATfE -16..+384 B; M4 -456..+312 B. Data/bss unchanged. Output/reference differences are unchanged by the pin and retained in the tables; no new numerical qualification or general int16 QUANTIZE correctness claim.
Measured candidate
900e41ce9f29cf1de682ca9d47e666b8546081f4/ CORE07fb2eeee1d74f9d257cc6ab5f643431a3bfb6ba, baselinee1f490191f23c83e45a799c2461b25af4149971f/ CORE94bf3b9362874d951f9363bb0ddef72fd162e8f6. Requested finalcf0490b1102ec145f5f0fbb66c511a863c1543de/ COREb8102da7ec38cf764230d6932bdc6e8b4e74474e: documented algorithm/compile-condition equivalence, not independently rebuilt final-ELF identity.Evidence:
/home/adamp/Ambiq/aitg/orchestration/artifacts/hpx/rt-426-ab/REPORT.md, all-model/per-layer/flag/size CSVs, image provenance, output audit, raw runs. Physical identity/claims, image hashes/full flash readbacks, dual reads and build IDs verified. An initial host J-Link crash produced no accepted result; preserved and unchanged capture completed. All claims released. No CI/publication/merge. Next: RT/LEAD disposition of layer flags or a bounded repeat assignment; board validation is complete at the stated measurement scope.apage224 commentedon Oct 6, 2026
Checkpoint: #426 board A/B complete; performance acceptance held.
Head remains cf0490b1102ec145f5f0fbb66c511a863c1543de, local/unpublished, base e1f4901. HPX completed186 captures /93fixed-input pairs, all full outputs byte-identical and finite. Measured candidate is RT900e41ce/CORE07fb2eee; requested final pin uses COREb8102da7. Source/compile-condition equivalence retained; no rebuilt final-head ELF identity claim.
Whole-model loss: M55 GCC max+0.47%, ATfE max+0.23%, M4 max0%. Layer criterion unresolved:49 initial positive flags exceed3% on exact integer cycles across26 board/compiler/model pairs. Of49,25 have either median>=1000cycles,24 are smaller;13 have zero within-capture spread. All49 remain OPEN; no independent repeats or noise/small-layer waivers. The1132→1166 ATfE CONV case is+3.0035% despite displayed3.00%, and stays flagged.
Next decision: authorize a bounded52-capture fresh B/A repeat over the26 affected pairs with the same retained images, or obtain Adam's explicit disposition of the listed costs/uncertainties under the layer policy. No new validation or CI performed. Existing publication+one-CI authorization was conditional on passing required A/B; the per-layer condition has not passed, so publication remains held.
Card/evidence: artifacts/rt/pin-ns-cmsis-nn-740/PUBLICATION-CARD.md, PERFORMANCE-DISPOSITION.md, LAYER-FLAG-DISPOSITIONS.csv, BOUNDED-REPEAT-PROPOSAL.json; raw HPX artifacts/hpx/rt-426-ab/REPORT.md and CSVs. Known supported-input s16 QUANTIZE limitation remains separate #427/#733; these corpus outputs do not cover it. No push/PR/CI/merge/release.
apage224 commentedon Oct 6, 2026
Plan for LEAD-authorized bounded follow-up: exactly four fresh A/B pairs (eight captures maximum), retained audited images unchanged. Apollo4 GCC arr_4_eff_sm_a8w8 A→B; Apollo510B GCC vww_a8w8 B→A; Apollo510B GCC vww_fp32_w16 A→B; Apollo510B ATfE arr_4_eff_sm_a8w8 B→A. Same warm LP96MHz, five warmups, nine model calls and fifteen layer calls; physical claim/identity/image-readback checks per capture. Stop on unexpected output/assertion or failed capture, without automatic retry.
Preserve the completed 186-image matrix, original tables/49 positive layer flags and raw spreads. Report repeat medians/spreads and whole-model changes separately, exact outputs, reproduced/unconfirmed selected flags, and untouched flags explicitly initial/unmeasured. Evidence will be under
/home/adamp/Ambiq/aitg/orchestration/artifacts/hpx/rt-426-repeats-20261006/. No rebuild, new threshold, additional matrix, CI or publication. Next checkpoint: the eight-capture result or a stop condition.apage224 commentedon Oct 6, 2026
The four authorized targeted A/B repeats completed in exactly eight captures, without retries or changes to retained images, compiler flags, configuration, placement or thresholds. Capture order alternated A→B/B→A. All four output pairs match each other and their original outputs; arena usage unchanged. Every physical claim, identity check, image readback, dual read and build ID passed; all claims released.
All 17 selected initial positive layer flags reproduced (>3%, same sign), with none unconfirmed and no new positive repeat flags. The remaining 32 positive layer flags are still initial/unmeasured, including IC int8 FC +18%. Reproduced primary layers:
Repeat whole-model changes: Apollo4 arr -6.607%; M55 GCC VWW int8 -0.194%; M55 GCC VWW FP32/FP16 weights -0.114%; M55 ATfE arr +0.009%. No positive whole-model repeat loss exceeds1%. Original and repeat medians/spreads are reported separately. The material layer costs remain explicit and require Adam disposition/final card before publication; no blanket performance PASS.
Evidence:
/home/adamp/Ambiq/aitg/orchestration/artifacts/hpx/rt-426-repeats-20261006/REPORT.md, totals/per-layer/selected-flag/untouched-flag CSVs, capture order, image/output provenance, raw captures and fresh-capture integrity hashes. Initial186-image artifacts, tables, report and results were hash-checked unchanged. Retained candidate source is RT900e41ce9f29cf1de682ca9d47e666b8546081f4/ CORE07fb2eeee1d74f9d257cc6ab5f643431a3bfb6ba; requested finalcf0490b1102ec145f5f0fbb66c511a863c1543dehas documented algorithm/compile-condition equivalence, without an exact-final ELF identity claim.No additional captures, matrix, CI or publication. Next: RT/LEAD/Adam disposition of the reproduced material-layer costs and remaining initial flags.
apage224 commentedon Oct 6, 2026
Checkpoint: all49 layer flags have contextual dispositions; staged repeat budget fixed.
RT final head unchanged cf0490b1102ec145f5f0fbb66c511a863c1543de, local/unpublished. Original186 captures/93 output-identical pairs accepted; whole-model loss bound passes, layer acceptance held. All49 remain OPEN, with raw-cycle selection unchanged.
LEAD authorized exactly five named HPX pairs/ten captures total inclusive: Apollo4 GCC arr int8, M55 GCC vww int8, M55 GCC vww fp32_w16, M55 ATfE arr int8, plus M55 GCC vww fp16-native. These selected pairs cover19 initial flags;30 unselected remain initial-only/unconfirmed. The earlier26pair/52capture proposal was NOT APPROVED. No extra requests or product changes.
All49 rows now include absolute cycles/spreads, whole-model impact, dtype/shapes, source-selected adapter entry, unchanged A/B RT source blobs, compiler/defines, placement and linked image delta. Static source context is not an actual leaf trace. Seventeen flags lie in unchanged local RESHAPE/float16 widening code; no source cause or performance waiver is claimed.
Remaining/next: await one HPX repeat report, update the single LAYER-COST-CARD.md and LAYER-COST-DISPOSITION.csv with repeated/unselected status, then return exact-head recommendation to LEAD/Adam. Material actual-layer costs require Adam disposition. No publication/CI/merge authorized by this analysis scope; #427 remains separate repaired-dependency hold. No new board captures or builds performed by RT.
apage224 commentedon Oct 6, 2026
LEAD has superseded the four-pair budget with five pairs / ten captures total, inclusive of the eight completed captures. Preserve that four-pair checkpoint; add only one retained Apollo510B/GCC native-FP16 VWW pair targeting CONV14 (+38486 cycles/+3.53% initially) and all other layers in the image. B→A order from the supplied fifth-pair plan; original four capture orders unchanged. Same firmware/compiler/configuration/placement and assertions, normal physical claims, no automatic retry. Consolidate all ten captures and all initial flags into one report, then hold. No26-pair matrix, rebuild, CI or publication.
apage224 commentedon Oct 6, 2026
Combined bounded repeat report complete: five pairs / ten captures total, inclusive of the original eight; only the authorized native-FP16 VWW pair was added, in B→A order. No retries, rebuilds or configuration/threshold changes. All five output pairs match each other and their original outputs; arena usage unchanged. Claims released.
19/19 selected original positive layer flags reproduce; 30 remain initial/unmeasured. The added M55/GCC native-FP16 VWW CONV14 repeats at1091213→1129622 cycles (+38409, +3.5198%; A/B spreads2734/1882); SOFTMAX29 repeats508→525 (+3.3465%, zero spreads). The original four-pair outcomes are preserved unchanged, including arr M4 conv5/13/21, VWW int8 conv26/FC28, FP16-weight dequantize36 and ATfE arr MUL/FC flags.
One new single-repeat positive flag: native-FP16 VWW DEPTHWISE21 moved from initial+2.6849% (55645→57139) to repeat+3.4972% (55645→57591, spreads1367/1513). It is not a reproduced original flag. Three original improvements (DEPTHWISE13/25, AVERAGE_POOL27 in the added pair) are unconfirmed because repeat improvements fall below3%; their original values remain listed. No selected result was deleted.
No positive whole-model repeat loss exceeds1%: M4 arr -6.607%; M55 GCC VWW int8 -0.194%, FP32/FP16 weights -0.114%, native FP16 -0.0268%; M55 ATfE arr +0.009%. Raw original/repeat medians and spreads are separate in the tables. These model totals do not waive the reproduced layer costs or untouched flags.
Evidence:
/home/adamp/Ambiq/aitg/orchestration/artifacts/hpx/rt-426-repeats-20261006/REPORT.md, all-original-positive-flags.csv (all49), new-repeat-flags.csv, totals/per-layer/selected/untouched CSVs, capture-order/image/output/capture-integrity records, and raw runs. The completed four-pair checkpoint is archived in four-pair-checkpoint/; its eight raw records/blobs and order entries are unchanged. The original186-image matrix/report/tables/results are hash-verified unchanged. Actual retained900e41ce/07fb2eee source and firmware identities preserved; no exact-finalcf0490b1 ELF claim.This completes the ten-capture ceiling. No further captures, 26-pair matrix, CI or publication. Hold for RT/LEAD/Adam material-cost disposition and final card; no blanket performance PASS.
apage224 commentedon Oct 6, 2026
Completed bounded performance disposition for head
cf0490b1102ec145f5f0fbb66c511a863c1543de(basee1f490191f23c83e45a799c2461b25af4149971f). Local implementation/reviews remain accepted; publication and CI held.HPX original 186 captures / 93 fixed-input output-identical pairs and whole-model loss bound remain accepted. Approved repeat budget completed: exactly five pairs / ten fresh captures inclusive, no retries/rebuilds. All five outputs match originals, arenas unchanged, repeat whole-model losses <1%; claims released. Measured candidate remains RT
900e41ce9f29cf1de682ca9d47e666b8546081f4/ CORE07fb2eeee1d74f9d257cc6ab5f643431a3bfb6ba, source/compile-condition equivalent to the requested final pin; no independently rebuilt exact-final ELF execution claim.All49 original >3% layer flags retained: 19 reproduced;30 unselected/initial-only/unconfirmed, none waived. Largest repeated absolute cost: M55/GCC native-FP16 VWW CONV14
1091213→1129622cycles (+38409/+3.5198%, spreads2734/1882). M4 arr CONV5/13/21 repeat +9.566/+7.126/+5.348% with zero spreads; M55/GCC VWW int8 CONV26 repeats +14740cycles/+10.157%. Untouched IC int8 FC +18% remains initial-only. Raw-cycle boundary1132→1166remains +3.0035% despite printed3.00%.One additional single-repeat uncertainty: native-FP16 VWW DEPTHWISE21 initial
55645→57139(+2.6849%, spreads1367/1060), repeat55645→57591(+1946/+3.4972%, spreads1367/1513). Kept separate from the49 original flags. Three initial improvements in the fifth pair are unconfirmed by its repeat and remain listed. Whole-model improvements and numerical parity do not clear layer costs.One authoritative cost card:
/home/adamp/Ambiq/aitg/orchestration/artifacts/rt/pin-ns-cmsis-nn-740/LAYER-COST-CARD.md. All49 rows include initial/repeated absolute cycles/spreads, whole-model cycles/spreads/impact, dtype/shapes, source-selected adapter/backend, compile/link/placement context inLAYER-COST-DISPOSITION.csvandCONTEXTUAL-LAYER-DISPOSITIONS.json; new observation and accepted source hashes recorded separately. RT adapter sources are unchanged;17 original flags are local RESHAPE/float16 widening rather than direct CORE compute. Selected CORE leaf/fallback/alias addresses were not traced; no causal attribution or noise waiver inferred.Recommendation: park #426 at the exact final head, retain current v7.39.3 pin, no more captures or CI, pending Adam's material-cost/uncertainty disposition or separate repaired-pin scope. This defers v7.40 fixes/improvements; retained baseline has the separate known s16 defect, so parking is not general int16 correctness acceptance. A repaired7.41 pin needs separate evidence/scope. Any later Prepare/static-check optimization needs actual attribution and a bounded CORE+RT consumer change; no immediate unchecked generic-conv switch exists in the inspected headers.
Remaining: Adam exact-head decision on reproduced costs and uncertainties, or subsequent concrete scope. Earlier26pair/52captureproposal NOTAPPROVED/superseded. No new product changes, validation, publication, CI, merge or board expansion. Independent #427 repaired-dependency hold retained. Next: preserve all work/evidence and await disposition.
Final exact-head publication card explicitly records Adam's two cost-disposition choices: accept19 reproduced costs +30 untested original costs/uncertainties +1 new single-repeat observation, or keep #426 parked/currentpin (recommended). Cost acceptance does not approve merge or blanket numerical/performance PASS. Publication/CI stay held; after disposition LEAD considers existing conditional publication/oneCI with normal head/main/claim/workflow checks and actual released7.40 CI assets. No further captures, builds or review rounds. PRIVATE PR-BODY and PERFORMANCE-DISPOSITION are synchronized to the same evidence/counts.
apage224 commentedon Oct 7, 2026
RT delivery/regression — one consumer and gate proposal
Recorded 2026-10-07T02:19:42Z. Source preflight complete; real release/pin and final consumer checks remain gated. No runtime source/pin edit, test/inference, build/capture, review allocation, publication/CI or denied transport/review retry performed. Local heads426cf049/42771d9 and old49layer flags remain preserved. This package proposes later checks; it activates none.
Actual adoption
Current RT main e1f4901 pins released CORE7.39.3 /94bf3b9362874d951f9363bb0ddef72fd162e8f6. Its helia conv/FC/quantize_common sources are byte-identical to retained RT426cf0490b1102ec145f5f0fbb66c511a863c1543de. No current adoption of737 or733 is claimed.
Smallest coherent later integration
Fixed input/model proposal
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 NPZ hashes/raw output hashes, source routes/hash records and exact427input/output/scales are in SOURCE-MODEL-PREFLIGHT.json, copied from HPX's existing frozen inventory, not regenerated. Older RT HC AD/VWW assets embed3d3b11/94daa4 models, different from selected9e6b99/8ba78d: HPX must fix packaging identities before any final image acceptance; old framework can be reused but old embedded data cannot qualify these rows.
Proposed RT contribution: four matched pairs above (eight images/once captures), plus at most one actualMVE QUANTIZE functional capsule reusing all three427cases. Maximum requested RT slots9images/9oncecaptures,0retries; no compiler/precision/cache/placement grid. HPX must size finite wall/artifact budget from existing workload costs before exact-image authority; none granted here. Capsule can be omitted only when equivalent exact-final MVE caller evidence already covers these cases, not because corpus lacks QUANTIZE. NoTCN without affected-route proof, no duplicateRTpacked/TCM AD sweep. Actual builds/captures require separate fixed-manifest approval.
Integer optimization acceptance is exact RTbefore/after byte parity. QUANTIZE repair uses independent existing expected saturated outputs rather than defective-baseline parity; preserve fail-before evidence. Record all model/operator deltas and ranges in cycles/µs/%; retain >1%model/>3%layer flags and material-cost decisions. Timed restore_invoke includes restoration once: no added restore totals, subtraction or sum-of-medians partition. Samples inside a capture are not independent repetitions.
Golden sources are known, not invented tolerances. NativeFP16golden is a FP32-sibling/cast reference with a different schedule; raw finite errors/task metrics must remain explicit. Representative labels/datasets and authoritative nativeFP16 application acceptance rule are missing from this packet. Fixed own-reference equality or one golden classification is not modelquality PASS. Prior LiteRT/AOT fiveoneLSB ResNet gap remains a separate numerical question, not a reason to relax integer parity.
Unresolved gates / preserved decisions
LEAD/operator: legitimate SSH restoration and737review resolution. CORE/Adam: design/cost, perPR merge and real containing release. RT/LEAD: new exact pin/head/consumer sequencing and publicationCI authority. HPX: selected-asset packaging, finalroute/fit/stack/readback manifests and finite budget. Modelowners/Adam: authoritative quality policy/task data. All named gaps are blockers to adoption acceptance, not blockers to this completed preflight.
RT426's49original cost flags (19reproduced/30initial)+1newseparate and exact-head/material-cost/publication holds remain; four proposed model pairs do not dispose those flags or qualify generalint16. No parked735/Q31/736 activation, source tuning, review retry, cleanup/release/pin/merge performed. Deeper cleanup/V1 decision follows delivery acceptance under separate scope.
Refs #426; Refs #427.
HPX dependency handoff:
/home/adamp/Ambiq/aitg/orchestration/artifacts/rt/delivery-regression-20261007/HPX-DEPENDENCY-HANDOFF.mdandRT427-RUNTIME-FIXTURE.jsonselect the existing N9 actual runtime QUANTIZE body/tail fixture with18Binput/18Boutput+2Btrailingcanary and source-declared exact saturation oracle. Source fixture only, no emitted model or execution. Selected Tiny routes/precisions, unchanged RTADstandard route, matched release/pin prerequisites, oldasset mismatch and absent task-quality criteria are explicit. Existing host-only scope and all gates unchanged.apage224 commentedon Oct 7, 2026
RT426/427 — one coherent local released7.41 package
Base independently resolved currentmain5ea9c80b84c2c02f8237c14c77c5529de4f88a1b (only three official-footer/docs files differ from earlier e1f490). Preserve all oldheads/cf049 and71d9/WIP/49originalflags19reproduced30initial+1new; no changes to foreign footer files. New owned branch rt/426-released741-s16/worktree helia-rt-426-released741.
Use actual accepted COREv7.41.0 source48ddbf7d863745eee108c68726f42fc82dabe0af/treec93b26fde16198204825f23f88f4e3912b35ea2e and canonical released source pack/headers; SDK compilers GNU14.2/ATfE19.1.5/armclang6.23.32 remain distinct from source-built local compiler. Update coherent Make dependency pin and existing external CMake/NSX provider requirements/owning docs and narrow fixtures. Carry only the registered RT427 N1/N9/N5 real QUANTIZE consumer/tests from71d9, no runtime kernel/router/packing change. Prefer one dependency+consumer commit/PR because tests depend on repaired pin. No new release/runtime version bump.
One actual M55MVE local consumer execution against real7.41 provider, including exact outputs/canary/actual API route identity/selective link proof; M4 execution only if support contract changes require it. Reuse saved wrong-baseline sensitivity and CORE release/compiler matrices, no broad grid. Normal hooks required. Inspect exact parent drift and compiler/header/provider identity, preserve actual toolchain labels. Prepare four-Tiny legalconsumer/control handoff using accepted portable canonical assets, with explicit no model inference/image and pending realconsumer/modelquality/cost gates.737/734 absent from7.41, not waiting for/incorporating them and no737gain claim; RT FP16 remains standard, not AOT packed.
Initial author effort <=40min from17:53Z; checkpoint exactgap before expansion. No push/PR/CI/image/board/capture/release/merge/reviewer allocation or denied-action retry. Return exact localhead/card and proportionate affected independent review/final PROPOSAL only. Refs #426; Refs #427.
apage224 commentedon Oct 7, 2026
Local author checkpoint: adopted the accepted CORE v7.41.0 release and carried the existing #427 s16 QUANTIZE regression in one coherent package.
RT head 637485f, branch rt/426-released741-s16, base 5ea9c80. CORE tag v7.41.0, source 48ddbf7d863745eee108c68726f42fc82dabe0af, tree c93b26fde16198204825f23f88f4e3912b35ea2e; actual release/asset acceptance reused. Local commit only; working tree clean.
Make pin/fallback, owning workflow/docs and external CMake/NSX canonical-header floor now require 7.41.0. Registered N1/N9/N5 consumer is byte-identical to preserved 71d93211431e7568444cf7c3956f9611d4c61579. The previously reviewed cf049 convolution-status query fixture is carried unchanged for this dependency; no production kernel/packing/dispatch source changed. Current main's three site/footer parent changes remain untouched.
Actual released-source M55/MVE FVP consumer passed 3/3 and M55 status fixture passed 24/24. M4-target non-MVE/DSP consumer passed 3/3 on the compatible Cortex-M55 Corstone FVP; no physical M4 qualification is claimed. Local source-built compiler ATfE 22.1.0 remains distinct from released SDK ATfE 19.1.5. Real/wrapped API, source/header/object/ELF identities and selective link behavior recorded. External real-SDK CMake/NSX floor objects compile; real old 7.39.3 headers reject. Historical baseline failure and missing-wide mutant evidence reused without reruns. Narrow source gates pass; normal commit used, no installed pre-commit hook claimed.
Exact local card: /home/adamp/Ambiq/aitg/orchestration/artifacts/rt/released741-local-20261007/LOCAL-AUTHOR-CARD.md. Canonical four-Tiny and separate legal QUANTIZE handoff: FINAL-CONSUMER-HANDOFF.md in the same directory. Independent successor review/final proposal is recorded but not executed.
Historical cf049/71d9 refs and all 49 original layer flags (19 reproduced, 30 initial) plus one separate new repeat flag remain preserved and open. RT AD stays standard/nonpacked. CORE #737/#734 are absent from 7.41.0; no ResNet gain claim. Model inference, final matched costs, task-quality acceptance, review, publication and CI remain pending. No push, PR, image, capture, merge, RT release, tolerance change or denied-review retry. Next: exact-head qualification/review disposition; parked after this bounded author handoff. Refs #427.
apage224 commentedon Oct 7, 2026
Corrective plan: close the one confirmed P2 pack/source-floor class at local RT head 637485f. Preserve that reviewed head with named ref rt/426-released741-reviewed-637; make one additive local correction.
Align the owning CMSIS-Pack generator/example and current pack/source guide floors with the compiled HELIA requirement. Keep historical include-export/capability facts distinct. Use the existing compiled guard as the floor authority for pack generation and independent advertised-floor validation; add a focused checker regression that rejects deliberately old advertised metadata even when XML and generator setting agree. Reuse the saved original generated PDSC and real-header rejection/acceptance probes as failure evidence; no runtime rerun or full pack/build matrix.
Pass criteria: generated actual manifests/PDSC/checker/example/current docs agree on 7.41; the metadata mutation is rejected; existing registration/pin/source gates and normal commit pass; runtime source/fixtures remain identical to reviewed 637. Return exact corrected card, PR draft and consumer handoff with original review provenance and one class closure proposal. Initial effort <=25 minutes. No push, PR, CI, model/FVP/board/image/capture, release, merge or denied retry. All 49 original cost flags plus one separate new repeat flag and task-quality gates retained. Refs #427.
apage224 commentedon Oct 7, 2026
helia-rt #431 — merged released s16 repair adoption and consumer regression
Completed integration scope CLOSED / MERGED. Adam approved exact source head and the recorded cost/coverage disposition. No further PR431 merge decision is pending. RT release, task-quality and broader product qualification remain unapproved.
Approved head 2fb38ef, source tree250ac34f50c314007bfade5b5ebb319fde0fbeb9. Actual merge 75b0be9, sole parent 57b0e1d, integrated tree 4c252051e66891fc85ab810da3276e420c0ff59f, merged at 2026-10-08T03:15:28Z. Current remote main at verification is the merge commit. Source branch rt/426-released741-s16 remains exact approved2fb, locally and remotely; no restoration, rebase or branch deletion occurred.
Independent owner verification 2026-10-08T03:19:27.089984+00:00: GitHub PR reports CLOSED/merged at the exact approved head and merge SHA; full written subject/body equal the guarded merge text. Owner independently read reviewed-base, approved-head, parent and merged Git trees, applied all20 approved changed paths to the actual parent, compared every2,396 leaf/path/mode/blob, and reconstructed Git tree SHA4c252051e66891fc85ab810da3276e420c0ff59f. Every approved changed blob/mode is exact2fb and every unrelated current-parent path is preserved.
Main advanced from the premerge card's2fd55fd56541a780a59b089e03be34b466f26efa to57b0e1d69b10ea630ab86c43a9aac4cdf4124617 through the separate branding432 commit. Root's bounded docs/build-consumer assessment is retained; independent owner tree comparison confirms preservation of that branding. Integrated tree correctly differs from approved-head tree. No fresh combined browser/render/runtime/CI run is claimed. The written conventional merge message retains Refs426/427 without attribution or automatic issue-closing wording; full message is in INDEPENDENT-MERGE-PROOF.json.
Adam cost/coverage disposition: Approval covers merging this exact repaired-dependency/regression package with its two VWW node29 softmax costs and unmeasured broader-portfolio risk. Current flags remain explicitly >3%: cold603→648cycles /6.281250→6.750000us (+45 /+0.468750us /+7.462687%, disjoint ranges); warm623→648 /6.489583→6.750000us (+25 /+0.260417us /+4.012841%, overlap). This is accepted merge cost, not threshold relaxation, statistical significance, noise clearance or kernel-causal proof.
At Apollo510B96MHz, matched ATfE22.1 production asm1/no-fast-math, application-O3/RT-Os/adapters-O2/CORE-O3, MRAM model/input and shared-SRAM arena, all eight cold/warm whole-model median costs increased below1%. Largest ResNet warm87.620458→87.961427ms (+340.968750us /+0.389143%). VWW cold149.690115→149.732729ms (+42.614583us /+0.028469%), warm149.601177→149.650479ms (+49.302083us /+0.032956%). Restore+Invoke already includes one copy; no restore-only double count or layer-median sum. All110 remaining operator rows/65 positive deltas plus KWS increases and complete model ranges remain in the accepted saved tables. There is no speedup claim.
Historical49 flags (19 reproduced /30 initial-only) and the separate new single-repeat observation remain retained broader-portfolio warnings. Their old RT e1f→900e /CORE94bf→07fb/7.40, model hashes, compiler/effective-Ofast/M4-M55/placement tuples differ from current exact2fb/7.41 canonical four. None of49 flagged model hashes matches the current canonical assets. Adam accepted the uncertainty for this merge; these old observations do not become new-head measurements, disappear, or qualify broad portfolio performance. Full mapping/current cost/source-placement notes remain at /home/adamp/Ambiq/aitg/orchestration/artifacts/rt/rt431-four-model-decision-20261008 and historical /home/adamp/Ambiq/aitg/orchestration/artifacts/rt/pin-ns-cmsis-nn-740/LAYER-COST-CARD.md. No new repeat/tuning request.
Actual arenas A=B: AD3,456B, ResNet106,868B, VWW176,964B, KWS45,620B. Linked image changes AD0B, ResNet+144B, VWW/KWS+368B; model/input addresses shift correspondingly for changed images. Static fit and runtime allocation/guards are accepted for these exact images; stack reservation49,152B is not high-water measurement. RT softmax adapter/CORE softmax C source match across pair, support-header/full-object/link-placement differ; executed mathematical leaf/LUT/scratch/cache/instruction causality unmeasured.
Remaining unqualified scopes: AD remains standard/nonpacked RT FP16, not AOT precise; stored FP32-sibling/cast reference is nonauthoritative, its max0.6484375 discrepancy at630/640 values unchanged, no tolerance invented or task-quality PASS. VWW is full canonical A16W8, not true s8. Task labels/split/metric/accepted threshold absent, stack highwater/live math-leaf trace unknown. Four graphs contain no QUANTIZE; actual registered caller/provider evidence is separate from the inactive proposed serialized N9 capsule and general corpus acceptance. CORE737/734 absent. Separate acceleration/FP16/GELU tasks are not activated. Product/release/quality/performance/denial holds outside this completed exact merge remain; there is no general RT release approval.
Review and hygiene provenance: Full premerge decision/source independence and local gate disclosures preserved in PUBLICATION-CARD-BEFORE.md and /home/adamp/Ambiq/aitg/orchestration/artifacts/lead/rt741-delivery-review-20261007. Source-first began fresh with neutral full-diff brief/saved INITIAL before author framing; complementary contract/necessity contexts and exact affected/final records retained, original NEEDS_WORK labels unchanged. One shared rule/parser and bounded archive guards; no newly re-reviewed imported kernel design. Existing whole-file pack-script format mismatch remains NOT_PASS; no installed hook run, bypass or fixer delta invented. No new tests/source/fixtures/decoder/statistical/native/image build or hardware run in this closeout.
Feedback / public body: Owner's complete premerge refresh was2026-10-08T00:22:59Z, no material finding; body then aligned in one separately authorized metadata PATCH at00:40:46Z, byte-exact LEAD candidate SHA4ee88a97f0a721a684df67d86fbc3997c8d132903ca30fca26e4d1265e22d4e6. Root reports fresh0reviews/0inline/0discussion before guarded merge. Owner now independently verified merged PR identity without another feedback/CI exploration. All original KWS CLI failure/three-field artifact correction and raw data retained.
Owner closeout: Independently verified merge/content and all original ref/worktree/WIP states; close only completed431 integration and registered427 consumer delivery scope in owner records. Owning426/427 checkpoints updated in place; broader quality/cost evidence retained and no automatic GitHub issue-state closure implied. Retire only the temporary closeout claim, preserve every worktree/source/branch/artifact. One substantive owner proof then park with session running. No automatic main-CI observer or new pin/release/hardware authority.
Proofs: /home/adamp/Ambiq/aitg/orchestration/artifacts/rt/rt431-merged-owner-closeout-20261008/INDEPENDENT-MERGE-PROOF.json and INDEPENDENT-LOCAL-PRESERVATION.json. Root approval/merge/drift proof retained at /home/adamp/Ambiq/aitg/orchestration/artifacts/lead/approved-merges-20261008/rt431. One transient TLS read timeout was recovered with one same normal read retry; no permission denial/bypass occurred. Initial broader-ref/text-status/absent-sentinel comparison notes retained; normalization matches Root's original snapshot and revealed no actual work drift. Refs #426. Refs #427.
apage224 commentedon Oct 7, 2026
Sibling corrective plan for ONE continuing R1 pack/source-floor class at 0e4d270. Preserve that exact head under rt/426-released741-reviewed-0e4 and make an additive local commit.
Make the selected build_pack repo_root own the floor explicitly passed into build_pdsc. Remove import-time tool-checkout floor state. Use one small numeric-guard parser for selected-root and packaged-guard reads. Require checker source context explicitly: plain PDSC uses its staged directory or --repo-root; .pack uses its own packaged guard, without unrelated tool-root fallback. Check all existing generator/checker/CLI readers and retain default/backend/history fixes.
Metadata-only controls: current/default and old matching metadata, two distinct owned roots, selected-root generation/validation and minimal PDSC/guard ZIP reader. Demonstrate the original 0e4 mismatch and corrected rejection of wrong-tool-root metadata. No full archive build, deletion, runtime/compiler/FVP/model/board run or new assets. Preserve prior reports and whole-file baseline formatter limitation.
Return exact clean local head/card/diffs and reader/class dispositions; update existing checkpoint in place. No independent reviewer allocation, push, PR, CI, merge, release, denied retry or relaxation of 49+1/task-quality holds. <=20-minute bounded package. Refs #427.
apage224 commentedon Oct 7, 2026
Tiny continuing R1 plan at b18b3e2. Preserve exact b18 under rt/426-released741-reviewed-b18 and all earlier refs/reports. Modify only archive loader and focused test: count ZIP entries, require exactly one PDSC and exactly one canonical compiled guard, and read the unique ZipInfo entries. Reject multiple descriptors, duplicate descriptor filenames and duplicate/missing canonical guards without relying on entry order.
Reuse the saved reviewer three-control failure evidence. Add directly affected minimal ZIP negatives with clear ambiguity errors; retain valid single-descriptor behavior and all six existing source-root/metadata checks. One small metadata run plus applicable Ruff/source gates and normal additive commit. No full pack/example/compiler/runtime/model/target runs, version framework, directory deletion or review allocation. <=10 minutes; return exact clean head/card/diff/proof and update existing checkpoint in place. All 49+1/quality/cost/publication/merge/denial holds retained. Refs #427.
apage224 commentedon Oct 7, 2026
Exact reviewed-head publication plan under LEAD approval: publish 2fb38ef/tree 250ac34f50c314007bfade5b5ebb319fde0fbeb9, full reviewed base 5ea9c80. Original affected R1 closure, explicit source-readiness judgment and separate final light PASS accepted; requested reviewer models are not runtime-attested.
Actual main independently resolved as 2fd55fd. Its sole delta is nonoverlapping landing hero/UI dependency metadata; no source/build/pack/test/changed-guide correction required. Authorized exact head remains unchanged; this is not merge approval or combined-tree qualification. Existing branch PR search and remote branch search found none; current owning issue feedback read and saved.
Normal additive push to origin, then one directly READY PR so opened runs ordinary Tests Entry Point and cmake-ssot-smoke once; same-event docs/source companions permitted. Local PR body ordering nit fixed without source change. No draft/promotion, ci:run_full label, dispatch, rerun, duplicate PR, extra observer or watch/poll. RT sole observer: one run/head/attempt registration read, then ONE own-session spaced completion activation at least 20 minutes later. At completion read and disposition all posted feedback before card.
No numerical/compiler/FVP/model/board/full-pack reruns solely for publication. Original runtime provenance stays at 637; physical M4/vector N9/model-quality/ResNet737/packed AD claims remain excluded. Imported CORE release is a production dependency change. All 49 original plus one repeat flag and model/provider/fit/stack/performance/product/Adam exact-head merge gates remain. No RT release/merge authority. Refs #427.
apage224 commentedon Oct 7, 2026
Conditional library stage activated after the scheduled exact PR431 ordinary CI/full-feedback checkpoint: all three attempt1 workflows success, full feedback has no material source/head correction. Prepare exactly one true baseline/candidate M55 generic archive pair, no source edits: baseline RT5ea9c80b84c2c02f8237c14c77c5529de4f88a1b + real CORE7.39.3/94bf3b9362874d951f9363bb0ddef72fd162e8f6; candidate RT2fb38ef08574520c781be1bf8197c4c72e2f0d32 + real7.41/48ddbf7d863745eee108c68726f42fc82dabe0af. Fresh owned source/worktree/dependency/output paths; preserve all refs/WIP; no build_helia.sh in held trees, clean, downloads or force.
Both use existing ATfE22.1, cortex_m_generic/cortex-m55, helia, release_with_logs, production CMSIS_NN_USE_REQUANTIZE_INLINE_ASSEMBLY=1 and effective shipped Make optimization classes. Freeze source/header/dependency/compiler/flag/macros/archives/members/symbols/ABI closure. Original 637 test controls lacked explicit asm1 and used default build; execute only candidate registered three s16 cases and24status fixture once at changed asm1/release_with_logs configuration on existing FVP. No new matrix/Unity/model/hardware/CI. If source/fixture adaptation or missing dependency is required, stop with concrete dependency; no silent correction.
Finite deadline including closure: 2026-10-07T20:51:36+00:00. Exact first-attempt library builds only. Return one coherent CI completion plus immutable archive/include/provider locks and handoff to LEAD/HPX, with generic logging/timing/NSX/platform and remaining final link/fit/stack/readback/model limits explicit. All49+1/quality/cost/merge/release/denial holds retained; no images/captures. Refs #427.
apage224 commentedon Oct 7, 2026
Existing-evidence delivery plan: fold LEAD-accepted matched KWS diagnostic results into the current exact PR431/2fb card and canonical426/427; preserve initial CLI/artifact-correction history and all holds. Within one <=15-minute source-only stage, specify exact remaining AD FP16, ResNet A16W8 and actual canonical VWW precision/assets, static provider versus observed-route limits, restore/output/event/allocation/guard and unchanged-producer/decoder contracts. Produce one durable HPX/LEAD packet; propose six images/six captures only if all three pairs are legally represented, otherwise name concrete blockers. No tests, CI/full-feedback queries, source edits, numerical runs, native libraries, image builds, hardware, reruns or merge/release. Refs #427.
apage224 commentedon Oct 8, 2026
Decision-card plan at exact published PR431 head2fb38ef08574520c781be1bf8197c4c72e2f0d32: reuse LEAD-independent raw/control/arithmetic acceptance of all four canonical matched diagnostics and saved source/review/ordinaryCI. One complete paginated current PR/issue/inline/review feedback refresh plus exact PR metadata, no Actions/CI query; stop if material feedback/head correction is required. Consolidate new VWW softmax costs, actual historical49+1 applicability and task-quality/live-leaf/stack limits into one exact-PR Adam decision. No source/tests/statistical/decoder reruns/builds/hardware. <=10minutes, claim release/handoff/park. Refs #427.
fix: adopt ns-cmsis-nn 7.41.0 and cover s16 quantization