[amd-callumm]

Purpose

Pull upstream changes up to and including PR #44282 to fix Transformers v5 support.

The merge was done by cherry-picking individual commits starting from the last merge point, resolving conflicts as they arose, then validating the final result builds and passes a breadth of functionality tests (no new failures since before the merge), fixing any issues if present.

Minor performance regressions, if present, can be followed up on since this merge unblocks other important work.

Test Plan

Run similar merge verification tests to #983

Test Result

Correctness tests: no regressions

Attention performance tests: no new regressions since the last merge; a previous 22% qwen3.5_35b TRITON_ATTN is now back within tolerance.

[mgehre-amd]

What do you mean by "The merge was done by cherry-picking individual commits"? I assume you were doing a "git merge upstream/main" to ensure that the merge-base etc are correctly recorded?

[amd-callumm]

What do you mean by "The merge was done by cherry-picking individual commits"? I assume you were doing a "git merge upstream/main" to ensure that the merge-base etc are correctly recorded?

This is what I did for the prior merge, but when checking git merge-base upstream/main rocm/gfx11 again for this merge, the last shared baseline is apparently still at April 28. I thought that a merge would have updated this, but apparently not... In hindsight, I wonder if it's because I committed the previous merge PR to gfx11 with a squash-and-merge rather than a merge commit? IIRC this was the only method that allowed pushing the changes without further conflict resolution.

Regardless, this created a messy situation where trying to do another git merge upstream/main for this led to hundreds of files conflicting in complicated ways. The same goes for a rebase. Cherry-picking commits was far cleaner with only a few minor conflicts, and seemed like the only practical way forward without a huge time investment to attempt a proper rebase. Leaving it alone might create more challenges down the line, though...

[amd-callumm]

Created an empty merge commit to bump the merge-base. From examining a diff between this branch and the upstream commit merged, all the remaining changes appear to be ROCm-specific additions and changes, or fork-exclusive github action jobs left out intentionally.