Skip to content

v5.0.x: coll/ucc: enable asymmetric datatype check in UCC without OMPI fallback#13657

Open
QiaoK wants to merge 1 commit intoopen-mpi:v5.0.xfrom
QiaoK:ompi_ucc_dt_check_5.0
Open

v5.0.x: coll/ucc: enable asymmetric datatype check in UCC without OMPI fallback#13657
QiaoK wants to merge 1 commit intoopen-mpi:v5.0.xfrom
QiaoK:ompi_ucc_dt_check_5.0

Conversation

@QiaoK
Copy link
Copy Markdown
Contributor

@QiaoK QiaoK commented Jan 16, 2026

This change enables UCC to perform asymmetric datatype checks internally for gather, gatherv, scatter, and scatterv operations. This feature is only available in UCC 1.8 or greater.

The motivation for this change is to avoid scenarios where a subset of processes fallback to OMPI while other processes proceed to UCC when asymmetric datatypes are supplied. By letting UCC handle the datatype asymmetry check, UCC can signal all processes to fallback when necessary, ensuring consistent collective operation execution across all ranks.

@QiaoK QiaoK added this to the v5.0.10 milestone Jan 16, 2026
@QiaoK QiaoK requested a review from janjust January 16, 2026 15:25
@QiaoK QiaoK self-assigned this Jan 16, 2026
@QiaoK QiaoK added the bug label Jan 16, 2026
@QiaoK QiaoK force-pushed the ompi_ucc_dt_check_5.0 branch from ba56dfc to d626656 Compare January 17, 2026 03:39
@janjust janjust changed the title coll/ucc: enable asymmetric datatype check in UCC without OMPI fallback v5.0.x: coll/ucc: enable asymmetric datatype check in UCC without OMPI fallback Jan 17, 2026
Copy link
Copy Markdown
Member

@jsquyres jsquyres left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Is this not a cherry pick of #13656?

If it actually is a cherry-pick -- even if there was some back-porting necessary -- then this PR should not be marked as notacherrypick and the commit message should include the cherry pick message.

@jsquyres
Copy link
Copy Markdown
Member

Commit message was updated; thanks. I removed notacherrypick from the PR description.

That being said, we can't merge this PR (or #13658) until the main PR is merged (#13656). CI is supposed to enforce that, but apparently that check is not working properly.

@QiaoK QiaoK force-pushed the ompi_ucc_dt_check_5.0 branch from a3181a8 to a093f6f Compare January 24, 2026 02:38
@QiaoK QiaoK force-pushed the ompi_ucc_dt_check_5.0 branch from a093f6f to f737e38 Compare April 9, 2026 15:39
This change enables UCC to perform asymmetric datatype checks internally
for gather, gatherv, scatter, and scatterv operations. This feature is
only available in UCC 1.8 or greater.

The motivation for this change is to avoid scenarios where a subset of
processes fallback to OMPI while other processes proceed to UCC when
asymmetric datatypes are supplied. By letting UCC handle the datatype
asymmetry check, UCC can signal all processes to fallback when necessary,
ensuring consistent collective operation execution across all ranks.

Signed-off-by: Qiao Kang <qkang@nvidia.com>
@QiaoK QiaoK force-pushed the ompi_ucc_dt_check_5.0 branch from f737e38 to db6fbf3 Compare April 9, 2026 22:08
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants