Repository navigation
PMAT-3477: apr parity measures the Qwen3.5 hybrid on CPU vs CUDA (C14) — hybrid arm, measured threshold basis, K-quant resolver - #3534
Merged
Conversation
…ead of failing in the dense loader (C14, #3090) Pmat-Ticket: PMAT-3477 Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…vs-llama.cpp leg row is keyed as such, qwen3.5-0.8b/9b get a measured GPU basis (n=5 lambda, n=3 gx10, DP4A falsifier as the known-bad), and the resolver prefers the Q4_K_M twin Pmat-Ticket: PMAT-3477 Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
noahgift
enabled auto-merge
September 19, 2026 16:03
…ing too (DP4A through the recurrence, +0.179 to -0.765) Pmat-Ticket: PMAT-3477 Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
|
§13.11 rung 1 — quorum shadow verdict Shadow mode: this records a verdict and merges nothing. A refusal |
noahgift
added a commit
that referenced
this pull request
Sep 19, 2026
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
guyernest
pushed a commit
to guyernest/aprender
that referenced
this pull request
Sep 29, 2026
* release: 0.68.2 bump-version.sh 0.68.2 (every workspace, facades included; --check green) and the CHANGELOG [0.68.2] section, per docs/specifications/06x-release-schedule.md §4.2. After this merges, rel-0682-autopilot runs APR-RELEASE-001 §4 T-1..T-4 + close: deep (T-1, local), pre-publish dogfood, tag + release, clean-room.yml dispatched on the tag (T-3, run id recorded), assets by command, preflight, cascade (T-4, automated), install, host + installer receipts, close. Pmat-Ticket: PMAT-3477 Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * release: 0.68.2 CHANGELOG — Qwen3.5 on CUDA up to 27B (paiml#3517, paiml#3527) and every PR merged since v0.68.1 Pmat-Ticket: PMAT-3477 Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * evidence(ladder): 0.68.2 receipts measured from the bump tree on lambda (sm_89) and gx10 (sm_121) — every required rung green on both hosts; qwen3-8b optional red on lambda (paiml#3532) Pmat-Ticket: PMAT-3477 Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * contracts: arm ladder-measured + ladder-green shapes for 0.68.2; contracts.nt regenerated with the two 0.68.2 ladder receipts as focus nodes (ONT-4c1, paiml#3508) Pmat-Ticket: PMAT-3477 Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * CHANGELOG 0.68.2: rows for paiml#3533, paiml#3534 (merged since the cut) Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * CHANGELOG 0.68.2: row for paiml#3528 Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * CHANGELOG 0.68.2: rows for paiml#3526, paiml#3530 (merged-since-tag predicate now 0) Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> * CHANGELOG 0.68.2: row for paiml#3536 Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> --------- Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com> Co-authored-by: Alfredo Deza <alfredo@deza.pe>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
0.68.2 T-2 preflight at main e729859, row C14:
apr parityon Qwen3.5 failed with "Failed to create model: … Qwen3.5 hybrid" — #3517 madeparity_refusal_forstop refusing qwen35 (it has a GPU forward now) butparity_03.rs::runstill built the denseOwnedQuantizedModel, so the previous UNMEASURED-TOOL report became a FAIL.What lands
parity_hybrid.rs, include!-ed at the end of parity.rs): forhybrid_forward_handles(arch)the CPUQwen35Modeland the GPUQwen35CudaModelrun the same prompt position by position and feed the existingSpcMetrics/ verdict / JSON path untouched; a header-only quant refusal still refuses hybrids whose GGML type has no GPU GEMV kernel (IQ4_XS etc.).evidence/parity/thresholds.yaml): the CPU-vs-llama.cpp leg row (min_cosine deliberately absent, CPU parity reference for qwen35: build it from the GPU methodology, not from the PR it measures #3303) is keyedqwen3.5-0.8b@cpu-vs-llama.cpp— under the bare namejudge()reached it for the GPU record and failed closed on a number absent for a different reason.qwen3.5-0.8b/qwen3.5-9bget GPU-vs-CPU rows at 0.99 with a measured basis: lambda 0.998971 ×5 / 0.999254 ×5 (stdev 0), gx10 0.999176 ×3 / 0.999351 ×3 (peer session, same head), known-bad = the DP4A-through-recurrence falsifier (cosine 0.466 → −0.246, the standing test). Argmax is not part of the rule: 1–4 near-tie flips per record on both hosts at cosine ≥ 0.9994, the same class as the dense known-good records.check_model_parity.sh): among IQ4_XS / Q4_K_M / UD-IQ2_XXS sharing a logical name the K-quant twin is picked (sort order picked IQ4_XS and reported UNMEASURED-TOOL for a model it could measure). Self-test 26/26 incl. four new rows.Measured by the orchestrator (lambda, RTX 4090, this head)
cargo test -p apr-cli --lib --features cuda parity205 passed;cargo fmt --all -- --checkclean; clippy -D warnings clean.scripts/check_model_parity.sh --manifest: PASS qwen3.5-0.8b (78 positions, min cosine 0.9990), PASS qwen3.5-9b (0.9993), dense rows unchanged.keep-open: #3477 (epic closes at the tag)
keep-open: #3090 (closes at the tag with the receipts)
🤖 Generated with Claude Code