perf: reduce repeated Triton autotuning for L2Norm - #1

Open

Kamleecoder wants to merge 1 commit into

Ascend:core_r0.16.0from

Kamleecoder:fix/l2norm-autotune

Commits on Aug 26, 2026

修复 L2Norm Triton kernel 在动态 token 数场景下因输入 T/NB 变化反复触发 autotune，导致训练 iteration 耗时异常增长的问题。
Kamleecoder
committed
484b4e7