trunk/f3d603baf74daa6ac8fbf15feb37f0edb6ac9ad3: Avoid re-inferring DeepSeek recipes in the TMA heuristic (#197458)
- PyTorch 近 90 天出现 1271 次
- PyTorch 近 90 天第 1257 次发版
- 上一次:同一天稍早 · ciflow/trunk/197843: [INITIAL] Update
发生了什么
Human Note Simple fix for DeepSeek-style scaling. Agent note Prepared with AI assistance. At K=128, valid DeepSeek scale shapes [M, 1] and [1, 1] also satisfy the shape-only rowwise predicate. Re-inferring their recipes inside the already-selected main-loop heuristic therefore makes get_tile_size reject RowWise and compilation fails. Both supported main-loop recipes have 128-element K blocks. Use that invariant here…
摘要按规则整理自下方来源原文