← Back to events
ActiveAIRelease3.219

viable/strict/1789809976: [distributed] Shrink all-gather layout test tensors (#197639)

What happened

Three redistribution tests allocate 8192x6144 tensors to check view versus split/cat behavior. Use 32x24 instead; keep batch sizes, the four-rank mesh, shard placements, operation checks, and value assertions unchanged. The layout decision does not depend on tensor byte size. Test plan: CPU source build and targeted lint pass. All three four-rank Gloo tests pass: baseline 9.228s, patched 3.472s; post-build repetitio…

Summary assembled by rule from the sources below

Why it's spreading

Sources