ciflow/torchtitan/194833: [inductor] Fix loop-local load CSE lifetime (#194786)
- PyTorch 近 90 天出现 71 次
- PyTorch 近 90 天第 71 次发版
- 上一次：同一天稍早 · trunk/fcebff2edcedc4a973b6124659b9b7932db725ad: Fix the version gate on the free-threaded profiler warning (#194697)
发生了什么
Manual cherry-pick of 60598ed ( #194786 ) onto release/2.14. Triton load placement and CSE lifetime used separate heuristics. An indirect load emitted inside a loop could remain cached after loop closure and be reused by an epilogue even though its temporary did not dominate that use. Use the selected load buffer as the source of truth for whether a CSE value survives loop closure, and stop hoisting loads that carry…
摘要按规则整理自下方来源原文