trunk/34a42ca521c2bda062ff75603e8fa75b44e7e0af: [distributed] Synchronize peers in NCCL2 timeout tests (#197587)
- PyTorch: 1179 events in the last 90 days
- PyTorch: 1165th Release in the last 90 days
- Previous: earlier the same day · ciflow/inductor/197620: Update
What happened
Replace five 30-second peer sleeps with bounded subgroup-store waits. Rank zero releases peers after verifying the timeout or trace; synchronization stays off the deliberately blocked CUDA stream. Test plan: CUDA source build; targeted lint; five NCCL2 tests pass on four H100s (176.1s before, 56.2s after) and on two GPUs (50.0s). Pull Request resolved: #197587 Approved by: https://github.com/tushar00jain
Summary assembled by rule from the sources below