preflight: re-baseline the CUDA profile for b10630 (v0.33.2 fold) - #239
Merged
Conversation
The payload pin moved b10488 = 9d77fa172 -> b10630 = d222767c7 with the v0.33.2 fold, so payload_pin failed and every value it guards was suspect until re-measured. Re-measured first, updated second -- the same discipline as the b10488 re-baseline. No expected value was edited; only the identity moved. Full preflight on maxusai/ollama:sync-0.33.2 (0.33.2-dynres-5-g2b95b4a) returned PASS=18 / FAIL=1 with payload_pin as the sole failure, which is the pin doing its job. The ladders came back byte-identical to the b10488 rows -- nemotron_h_omni [266, 266, 578, 2306, 3270], gemma4 [1102 x 5], qwen35 [1034, 1034, 1034, 2306, 4082] -- with the same budgets and payload proofs, and nemotron's pinned budget still landing 3328 -> 3270. After the identity update: VERDICT PASS, 19 PASS / 2 SKIP (the skips are arches with no pinned-budget expectation, by design). 903 re-checked at b10630 by source inspection and still required: the defect is intact and #27044 is still open. Recorded that it must keep hitting only the ids call site, since the non-ids path correctly uses ne11. First CUDA run ever to exercise poison_probe, and the first anywhere with node-level corroboration. The meter (llama/compat/801) was enabled on the container and reported clean over 33 nodes, so the fp16-accumulate class is now asserted at v.blk.31.ffn_down itself rather than inferred from decoded text -- the chain from 801 through the #230 check works end to end. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
VERDICT PASS 19/2 on maxusai/ollama:sync-0.33.2 (0.33.2-dynres-5-g2b95b4a); payload_pin re-baselined to d222767c7 with provenance in #239. Remaining surfaces are unvalidated and the release table says so. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
glennneuber
added a commit
that referenced
this pull request
Aug 29, 2026
This was referenced Aug 29, 2026
glennneuber
added a commit
that referenced
this pull request
Sep 4, 2026
Re-measured first: the three arch ladders on the sync-0.33.3 canary are byte-identical to the b10630 rows (nemotron [266, 266, 578, 2306, 3270], gemma4 [1102 x5], qwen35 [1034, 1034, 1034, 2306, 4082]) -- the third consecutive pin bump with unchanged numbers. Only the identity moves; no expected value is edited, as in #239. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
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
1 participant
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.
The v0.33.2 fold moved the payload pin
b10488 = 9d77fa172→b10630 = d222767c7, sopayload_pinfailed and everything it guards was suspect until re-measured.Re-measured first, updated second — no expected value was edited, only the identity moved.
payload_pinthe sole failure — the pin doing its jobThe ladders came back byte-identical to the b10488 rows: nemotron
[266, 266, 578, 2306, 3270], gemma4[1102 × 5], qwen35[1034, 1034, 1034, 2306, 4082], same budgets and payload proofs, nemotron's pinned budget still3328 → 3270. The 2 skips are arches with no pinned-budget expectation, by design.Two firsts
poison_proberan on CUDA for the first time — our newest previous CUDA run predated it. And it ran with node-level corroboration: the meter (llama/compat/801) was enabled on the container and reported clean over 33 nodes. So the fp16-accumulate class is now asserted atv.blk.31.ffn_downitself rather than inferred from decoded text — the chain from #228 (meter) through #230 (corroboration) works end to end on a real build.903
Re-checked at b10630 by source inspection: the defect is intact and llama.cpp#27044 is still open, so the patch stays. Recorded in the profile that it must keep hitting only the ids call site — the non-ids path at
mmq.cu:137correctly usesne11because its body does too, and a careless regeneration that 'fixed' both would over-allocate the wrong one.