ciflow/b200/200995: Route sub-256B scatter rows to elementwise kernel

发生了什么

The vectorized/TMA scatter fast paths launch at least a warp per entry, so below 256-byte slices the per-entry overhead dominates and the elementwise kernel is faster (2.13 routing, verified by padded-stride measurements showing 3-13x wins at 16-128B rows). Floors fast_scatter_kernel_eligible (C++ aten path) and the cuteDSL TMA/vec-scatter conds at 256 slice bytes; 256B+ rows keep TMA. Extends test_scatter_add_large…

摘要按规则整理自下方来源原文

来源