Skip to content

ScaleRLT: retain IQuest eager update gate failures - #34

Closed
0z5a wants to merge 1 commit into
ops/scalerlt-cuda-correctnessfrom
ops/scalerlt-iquest-eager-diagnostic
Closed

0z5a wants to merge 1 commit into
ops/scalerlt-cuda-correctnessfrom
ops/scalerlt-iquest-eager-diagnostic

Conversation

@0z5a

@0z5a 0z5a commented Oct 6, 2026

Copy link
Copy Markdown
Owner

The IQuest batched replay still fails the strict Adam update gate after matching the official eager attention operator family. This records all sixteen configurations and thirty-two paired steps with the same tokens, targets, optimizer and numerical tolerances as the prior audit.

Outputs and full gradients pass all32 comparisons, while complete parameter/update checks pass26/32 versus24/32 for the previous SDPA profiles. Six failures remain, so the candidate is unqualified. The Markdown table, complete raw results, executed snapshots and immutable hashes preserve the negative result. No provider, GPU performance, official40B execution or reward-convergence claim is added.

Validation: original fixture functions and all input matrices match; every parameter, Adam moment and step counter is checked with nonzero updates. The numerical audit naturally exits1 after completing all32 steps. Stack base: #33; evidence files only.

Signed-off-by: 0z5a <dezhen.lu@student.uni-tuebingen.de>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant