viable/strict/1790375212: Fix int32 overflow in cdist CUDA backward kernel (#198452)
- PyTorch: 1504 events in the last 90 days
- PyTorch: 1485th Release in the last 90 days
- Previous: earlier the same day · viable/strict/1790371571
What happened
Fixes #128791 . The forward half of that issue was already fixed by #188006 ; this fixes the same crash in the backward kernel. I reviewed the diff and the new test, and the tests below were run on our H200 server. The description below was drafted with Claude Code and reviewed by me: The forward crash reported in #128791 was fixed by #188006 , which caps the cdist/pdist forward grid and grid-strides over the output…
Summary assembled by rule from the sources below