You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Great question. I think you've discovered a flaw in the current CI design... we weren't prepared to always fetch the pathfinder wheel from the main branch. So right now with the backport branch we do have two different sources...
As a bandaid, I can manually backport all changes in the pathfinder from the main branch. Later we can improve the CI design to fetch pathfinder from the main branch in the 12.9.x CI (just like how we fetch 12.9.x wheels from the backport branch in the main CI, but in the reversed direction). WDYT?
Because on main we did not add a proper logic in deciding "which branch am I on?" by the time we cut a v13.x backport branch when v14 is out this logic would fail.
We need to either fix this in the main branch, or figure out a new maintenance strategy before the next CUDA major comes (#1199).
Closing as superseded by #1199, per leofang's 2026-03-05 comment — the residual here is bindings branching strategy, and #2675 is in flight against it.
The concrete 12.9.x vs main pathfinder fetch asymmetry has been recorded on #1199 so it survives this closure.
Activity
leofang commentedon Aug 26, 2025
Closed by #904. Not sure why GH did not close it automatically.
leofang commentedon Mar 5, 2026
Ran into this and realized we need a proper fix before closing.
With #904 the two branches currently diverge in the way they fetch the pathfinder artifact:
12.9.x:cuda-python/.github/workflows/test-wheel-linux.yml
Lines 133 to 149 in 2b6609f
main:cuda-python/.github/workflows/test-wheel-linux.yml
Lines 117 to 121 in 4a58adb
Because on
mainwe did not add a proper logic in deciding "which branch am I on?" by the time we cut av13.xbackport branch when v14 is out this logic would fail.We need to either fix this in the main branch, or figure out a new maintenance strategy before the next CUDA major comes (#1199).
rwgk commentedon Mar 17, 2026
I removed the P0 label because this is in the backlog. (I want to keep the P0, P1, P2 labels in sync with what we're actually doing.)
rparolin commentedon Aug 20, 2026
Closing as superseded by #1199, per leofang's 2026-03-05 comment — the residual here is bindings branching strategy, and #2675 is in flight against it.
The concrete
12.9.xvsmainpathfinder fetch asymmetry has been recorded on #1199 so it survives this closure.