Skip UCCL nixlbench ASIO tests due to flakiness (#2055)
## What? Skip UCCL nixlbench test due to frequent issues on teardown Follow up for https://github.com/ai-dynamo/nixl/pull/2000 ## Why? See https://github.com/ai-dynamo/nixl/issues/1999 Summary from triage agent: The UCCL backend crashes on teardown (the benchmark row printed and both processes logged "Destroying Engine…/Engine destroyed", then task 0 segfaulted — a race/crash in UCCL engine cleanup, not a hang or timeout). This is the known-flaky UCCL nixlbench issue https://github.com/ai-dynamo/nixl/issues/1999. Commit https://github.com/ai-dynamo/nixl/pull/2000 attempted to mitigate it but only commented out the ETCD-based UCCL loop (test_nixlbench.sh lines 122-132); the ASIO-based UCCL loop (lines 95-103) remained active and is what ran and crashed. The pairwise ucx-v1.22.x variant (stage 351) passed, confirming the failure is a nondeterministic UCCL crash. (The extensive pin_thread_to_numa/selectNICs ERROR/WARN lines are benign noise from the CI node's NUMA topology, not the cause.) <!-- This is an auto-generated comment: release notes by coderabbit.ai --> ## Summary by CodeRabbit * **Tests** * Disabled the GPU-gated UCCL ASIO transfer benchmark tests due to CI flakiness. * Added a follow-up reference to track re-enabling the tests. <!-- end of auto-generated comment: release notes by coderabbit.ai --> Signed-off-by: Ovidiu Mara <ovidium@nvidia.com>
O
ovidiusm committed
f7653b1a9745a68794c18a191e8f18406283809e
Parent: e1186f2
Committed by GitHub <noreply@github.com>
on 8/6/2026, 4:27:24 PM