Fix pipelined transpose empty rank skips

[fallintoplace]

Summary

• Return early from pipelined all-to-all when there are no source ranks, avoiding performance-report access to an empty rank list after paired pipelined transfers clear the vectors.

Validation

• git diff --check

[romerojosh]

Hi @fallintoplace, thanks for yet another contribution! I had to use codex to get a grasp on what exactly this PR was attempting to fix as the PR comment is a bit vague and doesn't get to the meat of what these changes are solving. Just some feedback since I know that these PRs are being driven by codex or similar tool.

For future reference, this PR seems to be specifically addressing an out of bounds access that can happen with performance reporting in the pipelined backends due to this line

cuDecomp/include/internal/comm_routines.h

Line 459 in 3a8d786

if (handle->performance_report_enable && src_ranks[0] != self_rank) {

because the src_ranks and dst_ranks arrays are resized to 0 when paired like in the code here:

cuDecomp/include/internal/transpose.h

Lines 694 to 696 in 3a8d786

	// Skip alltoall, this transfer was paired with previous one.
	dst_ranks.resize(0);
	src_ranks.resize(0);

The quick return in cudecompAlltoallPipelined is sufficient to address this issue. The other changes should be removed as per the inline comments.

[romerojosh]

/build

[github-actions]

🚀 Build workflow triggered! View run

[github-actions]

✅ Build workflow passed! View run

[romerojosh]

LGTM!