pytorch
d78b3909 - Explicitly destory ProcessGroup in allgather_coalesced_async test (#65513)

Commit
3 years ago
Explicitly destory ProcessGroup in allgather_coalesced_async test (#65513) Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/65513 The error in #65231 means some child threads were destructed before joined. I added some trace and prints and found that, in the failed tests, all `assertEqual` are passed, but the `ProcessGroupGloo` destructor wasn't called in one of the process. It could be due to the only guarantee that Python makes is that garbage collection MAY happen before the program exits. This commit adds an explicit `destroy_process_group()` to alleviate the problem. cc pietern mrshenli pritamdamania87 zhaojuanmao satgera rohan-varma gqchen aazzolini osalpekar jiayisuse SciPioneer H-Huang gcramer23 Test Plan: Imported from OSS Reviewed By: rohan-varma Differential Revision: D31134174 Pulled By: mrshenli fbshipit-source-id: 2e42fe93d3f16ce34681b591afc15a6ac0b9fab6
Author
Parents
Loading