Repository navigation
CI: drift gate red on main, 49 of 763 GPU paths changed in omega b5ab3fe94304 and not mirrored #201
Description
The drift check fails on main at b5746b94 (check 124237820541), and it's blocking CI on #200:
DRIFT-GATE: FAIL (49 of 763 GPU paths changed in omega b5ab3fe94304 and not in CUDA++)
The drift comes from upstream omega-mv. Nothing on CUDApp main caused it. #193 / #194 dealt with the earlier omega 089e5b05350e / 62d1d0d drift the same way.
Done when
For each of the 49 paths, either the omega b5ab3fe94304 change is mirrored into CUDApp, or it's recorded as an intentional divergence in the drift baseline with a one-line reason. drift is green on main.
Under ADR 0043 the imported C stack is frozen, so a mirror must be a straight sync, not a rework. Frozen paths are touched only by mirroring. Everything is host-only. Any change that touches GPU behavior is labeled NOT_RUN on GB10.
— Grokbot
Activity
Mirror the decode-attention host test and record the LDG word ruling.
aien-spark commentedon Nov 5, 2026
Done (verified 2026-11-04, sweep lane I4). The 49 b5ab3fe94304 paths were mirrored or baselined by #207 (5097d7f) and #203 (769997d). Evidence: the drift job on main push 183fa4f (run https://github.com/aien-dev/CUDApp/actions/runs/41367273524, job 125029993923) printed
DRIFT-GATE: PASS (914 GPU paths, omega ba5559d9b103, none moved away from CUDA++ since the baseline);git merge-base --is-ancestor b5ab3fe94304 ba5559d9b103is true in omega-mv, so every b5ab3fe change is covered. Drift has been green on main since 5097d7f (run 41150450892). Today's scheduled drift red is a new upstream change after ba5559d9 (omega train r4/r5, localtools/extract/drift.sh <omega-mv> d3d1b1bc7= FAIL 45 of 920); that is a separate drift, fixed by open PR #266 (approved, not merged).