Skip to content

[DO NOT MERGE] Reproduce flaky AnnIVFFlatTestF_int8.AnnIVFFlat/2 (recall 0) - #2815

Closed
bdice wants to merge 6 commits into
NVIDIA:mainfrom
bdice:ci-repro-ivf-flat-int8-recall0
Closed

bdice wants to merge 6 commits into
NVIDIA:mainfrom
bdice:ci-repro-ivf-flat-int8-recall0

Conversation

@bdice

@bdice bdice commented Oct 10, 2026

Copy link
Copy Markdown
Contributor

Reproducer for an intermittent failure in NEIGHBORS_ANN_IVF_FLAT_TEST AnnIVFFlatTestF_{int8,uint8}.AnnIVFFlat/2 (dim=2, CosineExpanded): the filtered search returns recall 0, or duplicated indices. Seen mostly on 12.2.2, amd64, rockylinux8, v100, earliest-driver.

This PR reduces CI to the V100 and H100 C++ test jobs, repeats the test, and adds diagnostics on failure. Part of #2783 (workstream N).

@copy-pr-bot

copy-pr-bot Bot commented Oct 10, 2026

Copy link
Copy Markdown

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

@bdice

bdice commented Oct 10, 2026

Copy link
Copy Markdown
Contributor Author

/ok to test 8ed8cff

@bdice

bdice commented Oct 11, 2026

Copy link
Copy Markdown
Contributor Author

/ok to test ebfad2b

@bdice

bdice commented Oct 11, 2026

Copy link
Copy Markdown
Contributor Author

Closing: this was a reproducer/diagnostic PR for workstream N of #2783. The bisect showed that #2249 fixed the original flake; the zero-norm cosine guard is in #2828.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

Status: Done

Development

Successfully merging this pull request may close these issues.

1 participant