ciflow/inductor/200700: [inductor] Run very large short reductions outside combo kernels
- PyTorch: 2235 events in the last 90 days
- PyTorch: 2206th Release in the last 90 days
- Previous: earlier the same day · ciflow/torchtitan/200591
What happened
Summary: The combo kernel partitioner already gives a reduction its own kernel when numel * rnumel exceeds LARGE_NUMELS , but only for long reductions (rnumel > 2048). A large short reduction still shares a combo kernel with small ones and inherits their config. For an all-persistent combo, select_combo_heuristics takes the configs of the sub-kernel with the largest rnumel, which is usually a small one. In a product…
Summary assembled by rule from the sources below