Fix batched silhouette score reduction aliasing - #2422
Conversation
|
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. |
This comment has been minimized.
This comment has been minimized.
f8c9939 to
6ec7aeb
Compare
This comment has been minimized.
This comment has been minimized.
This comment has been minimized.
This comment has been minimized.
6ec7aeb to
d24aff2
Compare
This comment has been minimized.
This comment has been minimized.
d24aff2 to
77aaa23
Compare
This comment has been minimized.
This comment has been minimized.
|
The negative-control run behaved as expected. With the presumed fix reverted, This confirms that the stress test detects the unfixed behavior in the target environment. The next step is to restore the fix and rerun the same matrix; the A100/CUDA 12.2 test should then pass. The separate NN-descent recall failure on L4 appears unrelated. |
77aaa23 to
20bd669
Compare
This comment has been minimized.
This comment has been minimized.
|
@csadorf, if you sign your commits, CI will start automatically. |
|
/ok to test be4b2b6 |
Stores the row-wise minimum for batched silhouette scores in a separate device vector, preventing the reduction output from overwriting matrix elements that may still be read.
Adds
BatchedMatchesNonBatchedAcrossMetricsAndChunkSizescoverage for four distance metrics, three chunk sizes, and both default and stream-pool handles. The test uses seed170, which reproduced the bug during concurrent A100 testing. It remains deterministic, single-process coverage and is not guaranteed to reproduce the scheduling-sensitive race.Validation
Regression validation used the temporary
BATCHED_SILHOUETTE_CONCURRENCY_TESTadded inef922834, originallye20bfffcbefore the rebase. The temporary test has since been reverted.As a negative control, reverting the fix while retaining the test caused the A100/CUDA 12.2 CI job to fail: 12 of 512 worker seeds produced 22 batched/non-batched mismatches above the
1e-4tolerance. This confirms that the stress workload detects the unfixed reduction-aliasing behavior in the target environment.Closes #2432