Junjie Wang (PyTorch) 3106a33e41 [fr] Fix one error in analysis script when subPG world size is smaller than global size (#156156) ...

Summary: We run into an interesting case when we see so many mismatches while lot of mismatch turns out to be a fully match. The reason is that we use the dump ranks (which is from 0 to 79) to compare against the local pg ranks (0 to 7) this leads to false positive of mismatches. We can just check whether dump ranks contain all expected ranks or not, that should be sufficient.

Test Plan:
Test with the failed case with the script and we now see the correct behavior + new unit test case.

Rollback Plan:

Differential Revision: D76775373

Pull Request resolved: https://github.com/pytorch/pytorch/pull/156156
Approved by: https://github.com/VieEeEw

2025-06-17 21:17:58 +00:00

..

components

[fr] Fix one error in analysis script when subPG world size is smaller than global size (#156156)

2025-06-17 21:17:58 +00:00

fr_trace.py

[BE] fix typos in tools/ (#156082)

2025-06-17 19:25:50 +00:00