Closed
[Question|问题咨询]: [add_layer_norm] tiling UB 预算计算中常驻参数计算不合理 #7213
nio006Add on  Oct 22Closed at  17 days ago
nio006
nio006Collaborators
Oct 22 Create

问题描述

AddLayerNormTilingImpl(norm/add_layer_norm/op_host/add_layer_norm_tiling.cpp)计算 UB 预算时,gamma/beta/bias 常驻项统一乘了 bufferNum;
当前 bufferNum 硬编码为 1(tiling.cpp:608,kernel 模板 BUFFER_NUM = 1 默认值),无功能影响;若未来启用双缓冲(bufferNum=2),BetterUB/SPECIAL_REDUCE 分支将高估权重固定开销——方向保守(不会 UB 溢出),但导致 rowPerTime 偏小、单趟少算行,且公式误导读者以为权重参与流水轮转。
gamma/beta/bias 是循环不变量:Process() 开头加载一次(add_layer_norm_normal_special_reduce.h:146-154),全核所有迭代复用,不参与 CopyIn→Compute→CopyOut 轮转,双缓冲对其无意义。建议删除计算beta、gamma、bias的buffersize时,删除乘以bufferNum。

likedislike
nio006nio006Collaborators
Oct 22 Added label: question
yuning_chen
yuning_chenCollaborators
Oct 22 Comments:

初步分析

感谢反馈。

现象理解:AddLayerNorm 的 tiling 把 gamma、beta、bias 的常驻大小也乘了 bufferNum。当前 bufferNum 是 1,没有功能影响;如果以后改成 2,会高估固定开销。

当前判断:初步判断,待复现确认。

涉及范围:norm/add_layer_norm 的 UB 预算。

下一步:核对 Process 里这三项是否只在开头加载一次。确认不参与轮转后,预算里不要再乘 bufferNum。

likedislike
yuning_chenyuning_chenCollaborators
Oct 22 Assign wangpengbo26 as responsible person
yuning_chen
yuning_chenCollaborators
Oct 22 Comments:

您好,该问题已收到,当前问题 @liujie12345678 处理中

likedislike
yuning_chenyuning_chenCollaborators
Oct 22 Removed responsible person wangpengbo26
yuning_chenyuning_chenCollaborators
Oct 22 Assign liujie12345678 as responsible person
liujie12345678liujie12345678Collaborators
Oct 23 Assign liuyun_nj as responsible person
liuyun_njliuyun_njCollaborators
17 days ago changed custom state from 进行中 to 已完成
liuyun_njliuyun_njCollaborators
17 days ago Close issue