[psv666]

Purpose

Add Kev-4B's first eager native Rust/CUDA text worker for Choice, Noul and Score, aligned with the pinned official BF16 default service. This completes the implementation scope of #27; human review remains pending.

Depends on the independently reviewable export/FP32-head slice in #170 at 85119d093fa3cb6ae5244b1d5c7eb2619084d9a2. Review only the worker follow-up. Both PRs target upstream main; until #170 merges, this PR's Files changed also contains that prerequisite. Preserve #170 first in the merge order.

Summary

HTTP -> Kev prompt/token/readout processor -> SerialScheduler
     -> shared Qwen eager state/question execution
     -> unchanged FP32 pointer head
     -> whole-question calibration -> ordered typed response

The worker validates a pinned CUDA-merged export, performs real mixed-question warmup before listening, and supports the shared frontend. Model-local planning matches official state/branch padding and projection shapes. Optional Qwen/CUDA operators match Torch 2.8 reductions, BF16 rotary buffers, FP32 GDN state and masked SDPA. Temporary state belongs to one question; no cross-request prefix reuse is added.

Optional CUDA builds use pinned PyTorch/CUTLASS headers and ship their license notices. Serving has no Python/libtorch runtime dependency. Existing Qwen default/reference callers retain their dispatch and ABI. CPU diagnostic exports must be regenerated with CUDA merging for this worker's serving contract.

Review follow-up (2026-10-31)

Current head 9501edeae0d9f293dc954b9de09def57af2ae173 contains main 5fa5179d57f9267bb0d20162004d8ce26352aa9d and updated #170 85119d093fa3cb6ae5244b1d5c7eb2619084d9a2 as ancestors; no force-push or rewritten original commits. The worker-only comparison starts at that updated export/head SHA, not the earlier main-merge commit.

• Main's prefix allocation refactor conflicted with Kev's private temporary-state allocator. The resolution shares prefix_layout/prefix_bytes while retaining the public profile guards and private per-question allocation. The two new capture_fixed/continue_fixed callers explicitly pass no Kev service window. A pre-fix compile failure recorded both missing arguments; the integrated build now passes.
• CPU boundary tests now verify both new prefix APIs reject LoRA, FP32 GDN, reference and Kev service numerics before device calls. No CUDA kernel or Kev model/service.rs arithmetic changed from b7c94b3.
• The updated export tests run with explicit CPU merging. The synthetic full-path tests exposed and fixed the missing NumPy dependency, without changing the CUDA serving contract.
• Fresh checks: 403 workspace Rust tests pass / 94 ordinary-command ignores, 24 Python tests pass, 26 backend CPU tests pass, fmt, Clippy, release, strict docs, Ruff and ABI contract checks pass (existing Laya warning retained). A fresh ordinary sm_89 CUDA build passes 26 GPU kernel tests under gpu run; the one checkpoint-dependent kernel test is explicitly filtered out. The separately enabled pinned-head oracle passes its original tolerances (logit 3.12924385e-7, probability 2.98023224e-8).
• Evidence gap: the old temporary directory, checkpoint exports and v1-v5 raw archive were unavailable on recovery. Full-model BF16/HTTP, the optional Kev CUDA profile, and full-checkpoint prefix-cache requalification were not rerun on this merged head. The 5,662-question campaign and 99-operator evidence below are historical. No threshold was relaxed; missing raw captures cannot be replaced by these summaries.
• Merge coordination: a disposable integration probe against perf(open-jev): optimize 9B prefill while preserving native outputs #196 at 9152ee46b2f8a2589d64f8f193919d8916b96618 still conflicts in docs/architecture.md and src/models/qwen3_5/native/src/model.rs. Kev: add pinned CPU export and FP32 pointer-head validation #170 must merge first; whichever of perf(open-jev): optimize 9B prefill while preserving native outputs #196/Kev: add eager native Rust/CUDA text worker with BF16 service parity #197 merges second must resolve the shared executor and rerun shared numerical checks. This constraint is not marked resolved and unrelated perf(open-jev): optimize 9B prefill while preserving native outputs #196 code is not folded into this PR.

Both PRs remain draft. New-head 9501ede CI and Docs now pass: Rust, benchmarks, Kev exporter, CUDA contract and CUDA compilation are green; docs deploy is skipped as configured. CUDA compilation is not full-model GPU qualification. Prior CI results below retain their original SHA.

Required human contributor answers

Reused from the contributor's explicit answers in #170, supplied on 2026-10-25 (translated):

1. Outcome: Add native model support for System1-Omni users that agrees with official reference results and has reproducible acceptance.
2. Highest-leverage variable: Reference fidelity and reproducible native coverage.
3. How the work moves it: Reuse existing backends and independent small PRs to lower integration and review cost. The requested first Kev slice is 4B, merge plus pointer head.
4. Customer and actual need: System1-Omni users need native model support with reference-consistent results.

Is a PR needed? The contributor explicitly requested implementation PRs where appropriate and now requested the final self-review and push. The export/head and complete-worker slices remain separately reviewable.

Observed failure evidence: The earlier BF16 run had 139/3,393 questions above the probability gate (maximum 0.06492195), Score-255 error 0.16675543 and three near-boundary decision flips. The committed diagnosis records causes and fixes; original failures, layer traces and intermediate unsuccessful probes were retained in the original checksummed handoff; that raw archive is currently unavailable after the temporary-directory loss described below. No gates were relaxed and no failing sample was removed.

Test Plan

System1-Omni Version / Commit: fetched main 5fa5179d57f9267bb0d20162004d8ce26352aa9d; export/head 85119d093fa3cb6ae5244b1d5c7eb2619084d9a2; complete worker 9501edeae0d9f293dc954b9de09def57af2ae173. Current diff: 4,828 authored-code-file / 7,261 total changed lines, 73 files; worker-only diff from the updated #170 head: 3,965 / 5,423, 65 files. Authored-code counts include comments in .rs/.py/.cu/.cuh/.c/.h/.cpp/.sh, excluding data/config/docs/licenses. This received a full component/integration review. The requested two-slice boundary is preserved; processor geometry, shared numerical execution and calibration must work together to qualify the complete worker. Numerical repair, boundary tests and documentation are separate follow-up commits.

Official 2074cc44256200bb37540103fa4572901faa10f3; base 1001bb4d826a52d1f399e183466143f4da7b741b; adapter/head 139fdd94f1b6a6ad80cc15e08fcb99cac885a101; temperature 2.406050072164233. Torch 2.8.0+cu128, Transformers 5.17.0, PEFT 0.21.0; BF16 backbone, FP32 head, merged LoRA, default SDPA/Graph/cache enabled, optional FLA/causal-conv1d absent. FP32 controls are diagnostics, not the primary acceptance target.

Linux x86_64, lab L20X-labeled GPU, one GPU through gpu run, sequential requests with fixed ordering and mixed questions. CUDA build sm_89 plus compute_89 PTX; Torch reports capability 9.0 while nvidia-smi reports 8.9, so there is no separate sm_90 qualification claim.

Main commands:

cargo fmt --all --check
cargo clippy --workspace --locked --all-targets --features omni-laya/serve -- -D warnings
cargo test --workspace --locked --features omni-laya/serve
cargo build --workspace --release --locked --features omni-laya/serve
python -m unittest discover -s tests/kev -p 'test_*.py' -v
mkdocs build --strict

Build and serving recipe, acceptance protocol, boundary method and shared regressions. Exact GPU commands, library hashes and raw logs are in the author-local evidence/precheck-v5/; full reproduction is gpu run --gpus 1 --wait 2h --timeout 2h --note 'Kev BF16 reproduction' -- bash /tmp/system1-issue-27/reproduce-bf16.sh /absolute/fresh/output on the recorded host. That script was syntax-checked; its component campaigns were executed, not an additional monolithic whole-corpus rerun.

Test Result / Evidence

Check	Result	Unchanged gate
Fixed BF16 acceptance	5,484 requests / 5,662 questions; exact prompt/token/readout and usage	Exact processing
Probability max absolute error	0.0024709404	<=0.01
Score raw / returned-scalar max error	0.00000305218 / 0 levels	<=0.1 levels, as used in #110
Choice / Noul near-boundary coverage	28 / 36; no observed decision disagreement	>=10 each; report every near-boundary difference
Far-boundary decisions	All agree	Choice top-two margin >0.02; Noul distance from 0.5 >0.01
Labeled task accuracy	Official/native both 76.5625%, 896 tasks	Drop <=1 pp aggregate / 2 pp per type
Brier / normalized Score MAE change	+8.76751e-9 / -2.57951e-9	Increase <=0.01 / <=0.005
Independent pointer head	Logit max 3.12924385e-7; probability max 2.98023224e-8	Original #170 gates

The 2,269-question new Noul pool is frozen and deduplicated before inference; actual official unrounded probabilities select 33 new boundaries before native runs. Every candidate is retained and compared. All 36 Noul boundaries qualify again in a fresh official/native replay with zero within-implementation probability drift. Original worst probability/Score cases match embeddings, all 32 residual layers and final norm. FP32 controls remain separate, including native versus historical FP32 Score-255 difference 0.28167720.

Historical v5 precheck at b7c94b3 (not rerun in full at the updated head):

• 399 workspace Rust tests pass, 92 ordinary-command ignores; relevant GPU/head/processing checks run explicitly. 22 Python tests, fmt, Clippy, release, backend contract checks, Ruff and strict docs pass. Existing unrelated Laya ABI warning is retained.
• Both fresh ordinary and Kev CUDA builds pass 27 existing kernel, 3 prefix, 3 readout and 2 Graph-mode tests each. All 85 old cs1_* exports remain. No new mandatory Python/libtorch runtime dependencies.
• All 99 operator comparisons are bit exact, including pre-Kev library A/B and JEMM-width 5,120 reference norms. Shared-engine tests use the pinned Kev weights with the ordinary Qwen profile; they do not establish other models' task accuracy.
• Rebuilt Kev library reproduces the frozen 52-request / 75-question boundary/contract capture byte for byte; real warmup, 55 worker/frontend/native response comparisons and 8 error checks pass. All 2,269 new requests also pass exact CPU prompt/token/readout comparison.
• Independent export/head slice separately passes its seven Python tests, Rust/head, fmt, Clippy and strict docs checks.

The full 5,662-question result is the retained v4 campaign. Final precheck changes packaging, coverage and documentation, with functional model/kernel files hash-unchanged; it freshly reruns the shared-engine and boundary/HTTP checks above. Historical v1-v4 failures and captures are preserved, not relabeled as fresh runs. Committed fixtures have active tests/generators; raw logs, captures, tensors and weights stay outside Git.

Previous-head b7c94b3 CI passes Rust, benchmarks, Kev exporter tests, CUDA contract checks and CUDA compilation; the current Docs build also passes (deploy skipped as configured). CUDA compilation is not GPU inference validation.

Demo / evidence

Before: no qualified native Kev text worker. After: Choice/Noul/Score requests are served through native and frontend paths with the frozen checks above. The linked recipe and numerical/boundary reports contain reproduction instructions and actual measured results. The original checksummed evidence and PR text were delivered under /tmp/system1-issue-27, but those files are now unavailable. The restored handoff records current review fixes/checks and the missing-evidence boundary; it does not recreate the historical raw captures. No latency/throughput/memory comparison was performed, so performance charts/video are N/A.

Merge Danger

Door: two-way. Additive serving support; existing CPU exports need regeneration for native serving.

Blast Radius: Qwen. New numerical entry points are optional and selected by Kev; old symbols, default dispatch, prefix/Graph checks and legacy arithmetic pass. Actual Cua/Open-Jev/JEMM task accuracy, other GPU architectures/sizes, optional FLA, arbitrary official concurrent batching, long-input generality and performance remain unverified. Remote CI is separate from the reported local results. No universal model-quality claim.

Self-review

Full automated/agent-assisted component review is complete. Contributor/maintainer review remains pending; this PR stays draft. Human checkboxes are intentionally unchecked.

• I have reviewed the full diff and addressed the issues I found.
• I have checked that the change follows the project's architecture and stays focused on the stated purpose.
• I have run the checks appropriate to this change and reported commands, results, and anything I could not verify above.
• I have checked that the PR description, documentation, and any accuracy or performance claims match the implementation and available evidence.

[hsliuustc0106]

Draft PR review: #197 "Kev: add eager native Rust/CUDA text worker with BF16 service parity"

Reviewer pass: first review, pass102 (2026-10-31). Verdict: COMMENTED (draft PR — no approval or readiness implied). ## Status block

• Base: main @ 22b7717c9 (PR-declared and merge-base-confirmed). Head: b7c94b34a60f35a5b88438259c94dca8089d08a4 (matches origin/pr/197 and PR API headRefOid; rechecked this pass).
• Diff (two-dot git diff 22b7717c9..b7c94b34a, identical to the GitHub API numbers): 72 files, +6,980/−87 (total changed lines 7,067, matching the body). The briefing's "87 files +7016/−784" does not match the current PR state at this head; API and local git agree on 72/+6,980/−87.
• Draft state: DRAFT (API isDraft: true); the body explicitly keeps human checkboxes unchecked pending human review — correct for a draft; not filled by this review.
• Standing sync note: merge-base 22b7717c9 is an older main; current tip e573260cc. git merge-tree shows no content conflicts against the current tip, but a main sync before ready is required (CI revalidation at the new base).
• Author: psv666; head hosted on the psv666 fork (codex/issue-27-kev-native).

Skill / evidence provenance

• Skill: .agents/skills/system1-omni-review/SKILL.md, bytes read this pass, SHA-256 58be3bc60806c4c5af7038a0c6a80fadb94f3360d714556a3e965c8903dde108 (commit ef9279638). References read: docs/architecture.md, CONTRIBUTING.md, references/repository-contracts.md.
• Separation of sources: [inspected] = diff/tree read at pinned SHAs; [run] = commands executed locally; [CI] = Actions API at the exact head; [author] = body claims not independently verified.

Relationship to #170 (checked as requested)

• #170 "Kev: add pinned CPU export and FP32 pointer-head validation" is still OPEN and DRAFT, head 8f765fdbf9eb782ab6f9d731df359699b70398ba. [CI/API]
• That SHA is an ancestor of #197's head (git merge-base --is-ancestor → yes), and the body pins exactly that SHA as the dependency. git log 8f765fdb..b7c94b34 shows the 10 worker commits on top. The dependency pin is currently accurate.
• F-1 (P1, merge-order dependency): #197 embeds the entire #170 slice; #170 must merge first. Until then, #197's Files-changed double-counts the export slice, and any rebase of draft #170 invalidates the pinned SHA and requires a #197 rebase. The body declares this ("Preserve #170 first in the merge order") — recorded as a standing constraint, not a defect. Trigger: merging #197 first or #170 moving. Consequence: duplicated review surface; stale pin.
• The export manifest contract is shared and consistent: recipe/kev/export.py (from #170) was extended after 8f765fdb (+22 lines: merge_numerics {device, accumulation, torch} written into kev_export.json; CUDA-merge default with torch==2.8.0+cu128, TF32 off, FP32-matmul-highest; --merge-device cpu kept for diagnostics), and the worker's checkpoint.rs pins every manifest field via recipe/kev/pins.json (merge_numerics.device == "cuda"). Consequence: CPU-merged exports from the original #170 form are rejected with an explicit message — the body's "CPU diagnostic exports must be regenerated with CUDA merging" is enforced in code, not just documented. [inspected]

Template / body check

• Four post-#138 contributor answers: present, explicitly reused from the contributor's #170 answers (2026-10-25) with attribution, plus "Is a PR needed?" and observed failure evidence (the 139/3,393-question BF16 gate failures with max errors, before the fixes). [inspected]
• Checklist: present, four boxes unchecked with the statement "Human checkboxes are intentionally unchecked" — appropriate for a draft; the author's own self-review statement is separate above it.
• Headings follow the template (Purpose / Test Plan / System1-Omni Version / Commit: / Test Result / Demo / evidence with an explicit N/A reason for performance charts / Self-review).
• Large-code-change disclosures present: authored vs total counts (4,654 authored / 7,067 total, 72 files full; worker-only slice 3,920 authored / 5,367 total, 63 files), component/integration review statement, two-slice split rationale, per-area validation. My recomputation of the worker slice (76b764ae..b7c94b34) gives 64 files, +5,269/−139 — approximately consistent with the body's 63-file claim; the small delta is range-definition ambiguity, not a discrepancy flag.

Commit-by-commit review [inspected]

1. 4be35df4 (#170 slice) pinned CPU export + FP32 pointer head (+1,764).
2. 8f765fdb (#170 tip) validate every pinned LoRA projection target (+39/−2).
3. 76b764ae merge commit "Merge current main into Kev export and head preparation" — parents 4be35df4 + 22b7717c9. Hygiene: parent 2 is exactly the PR merge-base, so the merge is a clean bring-up to the review base; diff vs parent 2 contains only the branch's own 20 files; no evil-merge content found.
4. 0102113f eager native Rust CUDA text worker + parity harness (28 files, +1,995).
5. 57e239b6 reject invalid parity evidence and report boundary flips (gate-integrity tests).
6. 1d44e4f5 docs: eager numerical and serving acceptance limits.
7. 490d80d3 fix: reference model lifetime for lint (+3/−1).
8. 33e1848e qualify against default BF16 service + Unicode tokenization fix (13 files, +990).
9. b446d6ee docs: mark BF16 qualification as failing — the honest failure record before the fix; history preserves the failure arc rather than rewriting it.
10. 224fa3a8 test: verify rendered prompts across the frozen service corpus.
11. 0efa1d9f fix: align eager native numerics with the pinned BF16 service (27 files, +1,653 — the Torch28 operator variants).
12. d850153a test: qualify Noul boundaries using official service outputs.
13. b7c94b34 docs: record BF16 acceptance and shared Qwen regression evidence.

Linear story: build → discover mismatch → mark failing → fix (Unicode, numerics) → requalify → document. Good hygiene.

Contract review (Valen-native bar)

Export-as-load-contract, hash-enforced — yes [inspected src/models/kev/native/src/checkpoint.rs]

• recipe/kev/pins.json is compiled into the binary (include_str!) and every manifest field must equal the pin (unsupported Kev export field otherwise).
• Exact file-set equality (7 files, no extras), per-file SHA-256 recomputed from disk, pinned base config.json/tokenizer.json/tokenizer_config.json hashes cross-checked, pinned FP32 head hash hard-coded, head tensor shapes checked (q/k [256, 2560] = Kev-4B).
• Shadow protection: rejects an unlisted model.safetensors that would shadow the verified shards; weight-map entries must be model.language_model.* in the two pinned shards.
• Pins match the body's declared identity: revision 139fdd94, base 1001bb4d, reference 2074cc44, torch 2.8.0+cu128, BF16 backbone / FP32 head. All verified against pins.json and export.py.

Wire rejection corpus vs the reference [inspected + run + CI]

• Committed fixtures: tests/kev/data/unicode-processing.json (Unicode prompt/token/readout corpus), boundary-sources.json (33 frozen boundary question sources selected by actual official unrounded probabilities), head.json, planning.json, processing.json — maintained fixtures with generators, not one-off output. Raw 5,662-question captures stay outside Git (declared).
• Gate mechanics tests (tests/kev/test_service_acceptance.py, test_validate.py) enforce that failures cannot be rounded away: score can fail while all probabilities pass; raw-vs-wire score gates independent; gate changes invalidate discovery; missing/duplicate/reordered cases rejected; non-finite values rejected. Boundary flips are always reported; near-boundary coverage ≥10 per kind with far-boundary margins declared.
• CI-wired portion: new kev-export CI job (CPU torch 2.8.0, no GPU/weights) runs python -m unittest discover -s tests/kev -p 'test_*.py'; CUDA contract checks and CUDA compilation cover the backend textually/compile-only. The full corpus replay is a GPU-local, documented procedure (recipe/kev/service-acceptance.md, noul-boundary.md), not CI-executable — correct scoping, but the recorded-corpus byte-parity at any future head remains an author-run check. [CI + author]

Finisher byte-parity (CPython float repr) [inspected src/models/kev/native/src/postprocess.rs]

• rounded() implements Python round(float(x), 4) with ties-to-even via format!("{value:.4}") parsing (Rust formats the exact f64 decimal and rounds half-to-even — matching CPython's _Py_dg_dtoa behavior for this use), then serde_json shortest-roundtrip serialization matches repr for the rounded values.
• billing_json() reproduces json.dumps(..., ensure_ascii=True) including \uXXXX escapes and surrogate pairs, with DEL (0x7f) correctly escaped; answer objects are built in the official insertion order ("billing tokenization observes it").
• Calibration: pinned temperature 2.406050072164233 (from pins), FP32 softmax, finite/range/sum validations, typed Noul/Choice/Score answer construction with mode/confidence formulas mirroring the official interpretation. Corpus/gate tests exercise these paths weight-free.

BF16 rounding locations [inspected CUDA]

• service_readout.cu (new, compile-guarded by CS1_KEV_SDPA): CUTLASS mem_eff_attention BF16 kernel instantiation matching PyTorch 2.8 SDPA (BF16 Q/K/V/mask, FP32 output_accum), additive mask in BF16 with the documented visibility rule; other builds do not compile it, so no cross-profile drift.
• norm.cu: gated_rms_norm_kernel templated Torch28 — replicates Torch 2.8 reduction order (ascending shuffle step=1..16 vs legacy warp_sum) and element mapping (lane + 32*j vs lane*PER), with identical BF16 rounding points (round_bf16(v*inv), round_bf16(w*normed), to_bf16(scaled*silu)); the legacy export instantiates <false> — old arithmetic unchanged.
• gdn_fp32.cu: Torch28 template on prepare (element mapping + shuffle order), FP32 state, cs1_gdn_prefill_fp32_x with optional final-state output and zero-state null; context lifetime documented to outlive stream drain.
• ABI: all new symbols additive (ops.h "Optional Kev Torch 2.8 default-service profile"); Rust cuda.rs resolves them with ? — hard-required for the Kev profile (.context("Kev requires recipe/kev/build_cuda.sh")), optional for everyone else.

No silent fallbacks / readiness / ownership [inspected]

• Model::enable_service_readout_numerics hard-fails on missing symbols, wrong 4B geometry, or any graph/lora/reference/batched/packed state, before first forward. Eager is explicit (load_with_graph(dir, lib, false)). No fallback path exists.
• main.rs: load → engine.warmup() (real mixed Choice/Noul/Score inference, error aborts startup) → then bind/listen. Readiness-after-real-init satisfied.
• Serial admission (SerialScheduler, executor mutex, poison → error), permit retained until all submitted work drains including row failure/disconnect (executor.rs sync-on-both-paths), HTTP content-type 415, admission 503 + Retry-After, cancel-on-drop. Matches the runtime contract and frontend requirements at this scope (own binary; shared-frontend integration claimed via model/service.rs + 55 worker/frontend/native comparisons [author-reported]).

Cross-PR hygiene [run]

• F-2 (P2, coordination): #197 genuinely conflicts with sibling #196 — git merge-tree origin/pr/196 origin/pr/197 reports content conflicts in docs/architecture.md and src/models/qwen3_5/native/src/model.rs (both extend shared Qwen execution from the same base; both also touch norm.cu/ops.h/cuda.rs, which auto-merge but are semantically adjacent). Whoever merges second must do a real conflict resolution in the shared executor file; plan the order (#170 → then coordinate #196/#197) and re-run shared numerical checks after resolution.
• #197 × #198: clean. Both × current main tip: clean (0 conflict markers).

CI (observed) [CI]

Actions API filtered to exact head b7c94b34a..., event pull_request:

• ThinkFlowLab/system1-omni: CI run 40649603220 success (the run ID the body cites); Docs run 40649603196 success.
• psv666/system1-omni (fork): no workflow runs recorded at this SHA (empty result). The body's CI link points at the upstream/mirror run; recorded as observed.

Locally executed [run]

Environment: macOS arm64, no cargo/GPU; Python 3.12.8, pytest 8.4.1, mkdocs 1.6.1; local conda torch is broken (missing libtorch_cpu.dylib). Worktree at b7c94b34a, removed after.

• PYTHONPATH=. python -m pytest tests/benchmarks -q → 133 passed.
• python -m unittest discover -s tests/kev -p 'test_*.py' → 16 tests: 15 passed, 1 error — the single error is test_export failing at import because the local torch install is broken (environment issue, not PR code). In CI (green at this head) torch==2.8.0 CPU is installed and the same command passes.
• mkdocs build --strict → exit 0.
• Rust checks and all GPU/head comparisons: not executable locally; CI covers compile + weight-free Python; GPU acceptance numbers are author-reported.

Material unverified checks

• The 5,484-request / 5,662-question BF16 acceptance campaign, the 99 bit-exact operator comparisons, the byte-for-byte 52-request boundary/contract capture, and the 2,269-question fresh Noul replay are author-reported (raw evidence is local to the author: evidence/precheck-v5/, /tmp/system1-issue-27/handoff.md).
• CUDA compilation CI is not GPU inference validation (the body says so itself); sm_89 build vs "Torch reports 9.0" discrepancy is disclosed but unresolved hardware nuance.
• Shared model/service.rs changes to the default Qwen consumers: code inspection shows additive/optional symbols only; existing Qwen regression suites pass per author + CI, but no independent GPU recheck was possible in this pass.
• #170 may move; re-pin and rebase before merge (F-1).

[psv666]

In response to this review.

Thanks for the review. Follow-up head: 9501edeae0d9f293dc954b9de09def57af2ae173; main 5fa5179d57f9267bb0d20162004d8ce26352aa9d and refreshed #170 85119d093fa3cb6ae5244b1d5c7eb2619084d9a2 are both ancestors. Original commits are retained; no rebase/force-push.

• F-1: agree with the merge-order constraint, rather than treating it as an observed runtime P1. Kev: add pinned CPU export and FP32 pointer-head validation #170 remains the prerequisite and must merge first. Updated the pinned SHA and worker-only diff boundary; moving Kev: add pinned CPU export and FP32 pointer-head validation #170 has been propagated to this branch.
• Main sync: completed, including semantic integration. The current main introduces shared prefix allocation and fixed-prefix capture/continuation. Resolved the allocator conflict without dropping Kev's private per-question storage or public profile guards. Compilation then exposed two missing service-window arguments in the new callers; both now explicitly select the ordinary Qwen path. Added CPU tests proving these APIs reject incompatible LoRA/FP32/reference/Kev profiles before any device call.
• F-2: confirmed as an outstanding integration constraint. Probed perf(open-jev): optimize 9B prefill while preserving native outputs #196 9152ee46b2f8a2589d64f8f193919d8916b96618 in a disposable checkout: docs/architecture.md and src/models/qwen3_5/native/src/model.rs still conflict. No unrelated sibling changes were merged into this PR. After Kev: add pinned CPU export and FP32 pointer-head validation #170, the second of perf(open-jev): optimize 9B prefill while preserving native outputs #196/Kev: add eager native Rust/CUDA text worker with BF16 service parity #197 to merge needs a real shared-executor resolution and fresh numerical checks; this is not claimed resolved.
• Fresh validation: 403 workspace Rust tests pass / 94 ignored; 24 Kev Python and 26 backend CPU tests pass; fmt, Clippy, release, strict docs, Ruff and ABI checks pass. A fresh ordinary CUDA library passes 26 GPU kernel tests under gpu run (one checkpoint-dependent test explicitly excluded). The separately enabled, freshly hash-verified head test passes unchanged gates. No CUDA kernel or Kev segmented-execution arithmetic changed from the reviewed b7c94b3.
• Evidence limitation made explicit: the old temporary directory, exports and v1-v5 raw archive were unavailable on recovery. Full-checkpoint BF16/HTTP, optional Kev-profile and prefix-cache requalification are not fresh passes at this merged head. Historical metrics/gates remain in the committed reports, with an availability notice; they do not replace the missing raw artifacts. Both PRs stay draft, human checkboxes untouched.

The NumPy dependency fix and synthetic full-export tests are reviewed independently in #170 and inherited here; the worker tests explicitly request CPU merging, preserving the CUDA serving-export contract.

New-head 9501ede CI and Docs now pass: Rust, benchmarks, Kev exporter, CUDA contract and CUDA compilation are green; docs deploy is skipped as configured. CUDA compilation is not full-model GPU qualification.

[MrDongsls]

Verified at 9501ede. The worker follows the pinned reference numerics closely, the optional Qwen entry points check out at every call site, the scheduler and executor lifetimes are sound, and CI is green at this head. Two small findings.

P3 in recipe/kev/reference-requirements.txt:3 the bare torch==2.8.0 pin resolves from the default PyPI index while the qualification environment is torch 2.8.0+cu128, so a fresh reference install can silently drift from the CUDA build the native side compares against, state the cu128 index requirement in the file or pin the local version.

P3 in src/models/kev/native/src/postprocess.rs:27 the rounded helper formats a string and calls expect on the serving path, so a future formula change that yields NaN or inf turns a formatting failure into a panic instead of a controlled error, returning a Result and propagating it as a 4xx or 500 would be safer.

[psv666]

@MrDongsls Thanks for checking the current worker. I verified the two follow-ups separately.

• Reference wheel — fixed in 125251cdcb343b089108bb454269ef9ca686dbbd. reference-requirements.txt now requires torch==2.8.0+cu128 and includes the CUDA 12.8 wheel index. The native recipe gives the complete separate-environment installation command. A dependency-resolution dry run against the existing CPU environment selected 2.8.0+cu128 in place of 2.8.0+cpu; nothing was installed. Strict MkDocs and git diff --check pass. This pins setup; it is not fresh GPU qualification.
• rounded() panic — the stated trigger does not reproduce. Executing the current formatting/parse expression with NaN, +Inf and -Inf yields Ok(NaN), Ok(inf) and Ok(-inf), respectively; those values do not cause parse().expect() to panic. In the current serving path, answers() first rejects non-finite/out-of-range probabilities and an invalid sum, and the admitted candidate count is bounded. The derived values therefore stay finite. I retained the arithmetic and signatures rather than treating a hypothetical future formula change as a current defect. A future arithmetic change should preserve these preconditions or introduce explicit finite-result errors with its own regression.

This commit only changes the reference requirements and installation documentation. The existing missing historical raw-evidence and #196 integration limitations remain; the PR remains draft.

Current-head 125251c CI is now green: CI passed Rust, benchmarks, Kev exporter, CUDA contract and CUDA compilation; Docs passed (deployment skipped as configured). CUDA compilation is not GPU inference qualification.

[Levius-Fubuki]

2026-11-02 实测复核：PR #197 — 当前 SM120 配置暂缓推进

PR: #197
Head: 125251cdcb343b089108bb454269ef9ca686dbbd
Base: 5fa5179d57f9267bb0d20162004d8ce26352aa9d

推进阻塞：完整BF16默认服务数值门槛失败

本次使用 NVIDIA RTX6000D / SM120 /156SM、driver595.91.07、NVCC13.2.86，固定官方 Kev2074cc4、4B base1001bb4、adapter/head139fdd9；Torch2.8.0+cu128、Transformers5.17.0等参考环境按配方固定，FLA和causal-conv1d均未安装。导出使用CUDA FP32 merge、关闭TF32，head为发布版FP32。

在原协议的全部5484请求 /5662问题上：

配置	超过0.01概率误差门槛的问题	最大绝对概率误差	总体结果
系统cuBLASLt13.1.0（130100），原始构建	202	0.06253224611282349	FAIL
受控诊断：仅改为Torch wheel自带cuBLAS12.8.4.1	181	0.02785402536392212	FAIL

诊断保持相同PR、Rust二进制、GPU、导出、输入、官方原始输出、NVCC13.2及CPU Torch2.8头文件；重建匹配该链接配置的旧Qwen库作算子回归。未改源码或阈值，原始失败完整保留。更换cuBLAS不足以消除偏差，尚未定位到唯一根因；不能据此声称某个算子必定有bug。

需要在推进前解决：修复并提交满足原0.01概率门槛的完整模型实测；或明确约束并验证受支持的GPU/工具链配置，使本次SM120结果不被算作已通过。若仍以历史SM89/PTX配置为验收依据，需要提供当前head、固定环境和可保留原始输出的新回放。不能仅放宽门槛、使用舍入后的答案代替未舍入概率，或将99项算子通过等同完整模型通过。

相关入口：src/models/kev/native/src/executor.rs:49-53（加载模型并开启service readout numerics）；配方和硬件记录见recipe/kev/native.md及service-acceptance.md。问题是本次配置的整体验收失败，尚未将错误归因到该入口中的具体一行实现。

已通过的范围

• 独立Cargo target：workspace403 passed /0 failed /94 ignored；fmt、严格Clippy、release build通过。Python24 passed。
• 发布版真实tokenizer的token/readout/计费检查通过；全部5484条预处理逐项通过。
• 真实CUDA导出248个投影成功，完整输入及输出哈希固定。
• 两种链接配置各99项算子对比通过，包括旧Qwen入口回归。
• 固定语料1213+2002+2269在完整模型推理前冻结，来源为已固定的官方development分区，不读取locked test分区。
• 官方服务保持BF16/FP32 head、SDPA、默认Graph与prefix cache，未使用KEV_*覆盖；首组实际43次Graph捕获、2246次回放、30次prefix cache命中。
• 全部官方输出在native之前产生，边界成员仅依官方输出选择，得到Choice30、Noul34；没有根据native结果删样本。主数值门槛失败，因此未把它包装成最终边界资格通过。
• 两种链接配置均通过39条直接服务/前端/native响应一致性比较及8项HTTP错误检查。传输通过与数值失败分别记录。
• task-quality指标没有触发原协议门槛；这不能覆盖202/181项概率门槛失败。没有提出性能提升结论。

依赖与证据

先推进#170；#197包含该前置工作。与#196共享Qwen模型/CUDA修改，独立分支验证不等于冲突解决后的集成验收。

证据与 review 范围

以上是2026-11-02本轮独立运行的结果，不是转述作者历史 benchmark。发出前已复核本条绑定的 head/base 未变。所有 Rust 计数均来自各 PR 隔离 Cargo target 后的重建；初版共享 target 的 Rust 结果已作废，ignored 未计为通过。

沿用可信 base 的 .agents/skills/system1-omni-review/SKILL.md，源提交 5fa5179d57f9267bb0d20162004d8ce26352aa9d，文件 SHA256 3a4b876162da14eee8db56af86664576aa5755b8f3e80b3072e944018647c3bc；已读取 contributor/architecture、repository contracts、artifact hygiene 及 CI/Docs 规范。本条补充当前提交的实测证据，不代表逐条关闭旧 head 上的全部 review 事项。

下列证据包保留原始成功/失败结果及环境、命令和输入 hash，已本地归档并核验 SHA256，尚未上传为公开附件；文件名与 hash 用于标识本次记录，不是公开下载链接。仅发布 COMMENT review；不代填人工清单、不变更 Draft、不执行合并。

可定位的失败示例：初始构建的v4/transfer-v4:mmlu/test/4720 / answer，最大概率绝对误差0.022282272577285767，超过原0.01门槛。这是数值失败示例，不是将全部原始/wire翻转都误记为受保护决策翻转。

归档：

• 197-full-initial-evidence.tar.gz，SHA256 9c4c60c6d3244fcab6ec8e4a47bf0f59ad135e30a8a1dd7c21d969bdcad7b2d5。
• 197-diagnostic-evidence.tar.gz，SHA256 b31f967b7d3956db2eb53ccaefa4f1e36989c1b80228f240930187e2a253e283。

保留全部官方/native原始输出、comparison、固定语料、边界成员、软件与库版本、命令、hash及失败状态。离线重算确认5662条逐题结果中的202/181项仅触发probability失败，最大误差与上表一致。初始文字将SDK13.2标签用于cuBLAS，现已核实并勘误：系统实际130100（13.1.0），Torch wheel实际120804（12.8.4.1）；原始证据未改写。旧SM89/PTX历史通过与本次SM120失败不能互相替代，当前没有证明所有硬件配置均失败。