[ariscript]

test_all_gather_sdma_in_place wipes out and refills this rank's slice before each call, with no
barrier before the all-gather. A faster rank can then start the next call and copy into a slower
rank's out before that rank's wipe, which erases it, and the slower rank finds an empty slice (the
same hazard example 34 had). _gather_and_check already barriers at that point.

It failed this way in CI on #617 and #618 (4 runs, 3 of them at 8 ranks), leaving the remaining ranks
hung until the NCCL timeout.

🤖 Generated with Claude Code

[Copilot]

Warning

Copilot couldn't run its full agentic review because it didn't start before the timeout. Make sure your repository has a runner available, or add a copilot-code-review.yml file specifying one with the runs-on attribute. See the docs for more details.

Copilot review overview

No open issues 🎉

What changed in this PR

Adds synchronization to prevent a race in test_all_gather_sdma_in_place where faster ranks can start the next all-gather and overwrite a slower rank’s out buffer before it wipes/refills, leading to incorrect assertions and CI hangs/timeouts.

Changes:

• Insert a ctx.barrier() between refilling the local slice and invoking all_gather to force all ranks to finish wipe/refill before any rank begins the next collective.

File	Description
tests/​ccl/​test_all_gather_sdma.py	Adds a barrier to prevent cross-iteration buffer overwrite races in the in-place SDMA all-gather test.

_{Review effort: Lite}