PyTorch reverts CSE optimization after autorevert triggered
A performance-enhancing commit for common subexpression elimination got rolled back automatically.
Deep Dive
pytorchmergebot auto-reverted commit 99f3e8d on July 22, undoing PR #188180’s "Reuse full-range splits for CSE". The revert was triggered automatically by pytorch's autorevert system, which can be disabled by adding the `autorevert: disable` tag.
Key Points
- PyTorch auto-revert rolled back commit 99f3e8d (PR #188180) on July 22 due to stability issues.
- The reverted change optimized CSE by reusing full-range splits to reduce memory operations.
- Manual review is needed before re-submitting; nightly builds may see a temporary performance dip for some model graphs.
Why It Matters
Highlights the fragility of compiler optimizations in PyTorch and the importance of automated safeguards for production frameworks.