trunk/e2d110a75ba155e0699f6de302a055154551e460: [distributed] Stabilize NCCL timeout tests (#194954)
- PyTorch: 209 events in the last 90 days
- PyTorch: 206th Release in the last 90 days
- Previous: earlier the same day · trunk/e2dee2a4b9e91bfe6374bf39be9449bb438e08c5: Revert "Drop unused includes in c10 (#194959)"
What happened
The timeout-extension test could enqueue its second work after the first extended timeout elapsed when ranks started unevenly. The asymmetric FX-cache test applied its short timeout to the initial cache warmup even though only the second phase exercises asymmetric compilation. Synchronize before the timeout window and delay the short timeout until the asymmetric phase. Test Plan: env CUDA_VISIBLE_DEVICES=0,1 NCCL_DE…
Summary assembled by rule from the sources below