trunk/834c8b143d838a6a4a5c7bc8cb1aebcec51a087f: [distributed] Avoid duplicate file-rendezvous test runs (#198049)
- PyTorch: 1780 events in the last 90 days
- PyTorch: 1758th Release in the last 90 days
- Previous: earlier the same day · trunk/42c892a1eef5dd839d7bc8e64bd96dfebcfa1b0e: Comment NIXL transport lifecycle states (#198890)
What happened
test_distributed runs each backend twice via for with_init_file in {True, False} , but the loop only changes the report tag; both runs use the same init method and command. Run each backend once. Local: 6 NCCL and 6 Gloo cases took 205s before, 103s after. Test plan: CUDA build and lint; NCCL/Gloo collective and timeout tests and the full quantization suite through run_test.py. Pull Request resolved: #198049 Approve…
Summary assembled by rule from the sources below