ciflow/inductor-rocm-mi200/192524: [c10d] Scope the symmetric-memory lifecycle review fixes to ROCm
- PyTorch: 1070 events in the last 90 days
- PyTorch: 1057th Release in the last 90 days
- Previous: earlier the same day · trunk/20665f569c2701914aa26fc95db15a453c10fa2a: Revert "Add CPython test `test_dataclasses` (#195855)"
What happened
A review pass over the previous commit found residual defects in the symmetric-memory registry lifecycle. All of them share one root cause: ncclComm_t addresses are recycled. A process group's predecessor is destroyed before its successor is created, so the allocator routinely hands the successor the predecessor's address, and any bookkeeping that identifies a registration by pointer alone cannot tell the two apart.…
Summary assembled by rule from the sources below