Closed
[Bug-Report|缺陷反馈]: aclnnDeepNorm 未校验 meanOut/rstdOut 归约维(尾维)形状,非法 shape 两段返回 0/0 且越界写 #7296
liuhanwen8Add on 28 days agoClosed at 6 days ago
28 days ago Added label: bug-report
yuning_chen
26 days ago Comments:
26 days ago Comments:
初步分析
感谢反馈,我们已开始核对这个问题。
现象理解:缺陷反馈;aclnnDeepNorm 未校验 meanOut/rstdOut 归约维(尾维)形状,非法 shape 两段返回 0/0 且越界写。
当前判断:从描述看,这是缺陷反馈「aclnnDeepNorm 未校验 meanOut/rstdOut 归约维(尾维)形状,非法 shape 两段返回 0/0 且越界写」(正文涉及返回码 161002)。正文已给出对照用例与期望/实际差异,更像参数校验、错误码语义或静默错值类问题。以上为初步判断,待复现确认,不能当成已确认根因。
涉及范围:aclnnDeepNorm、aclnnDeepNormGrad、aclnnop
下一步:我们会对照资料契约与实现路径继续定位,并按需补充回归用例。


26 days ago Assign liujie12345678 as responsible person
yuning_chen
26 days ago Comments:
26 days ago Comments:
您好,问题已收到,当前 @liujie12345678 正在处理中。


24 days ago Assign liuyun_nj as responsible person
11 days ago Associated pull request
10 days ago Removed responsible person liujie12345678
6 days ago changed custom state from 进行中 to 已完成
6 days ago Close issue
3 days ago Assign chenjiayi_ as responsible person
chenjiayi_
3 days ago Comments:
3 days ago Comments:
补充说明本问题的后续处理:
已修复并随关联 PR 合入。
处理结果:
- !14221:fix issue,add md doc,add check shape
验证:
- 关联 PR(2026-11-13)已合入,变更可从 PR 记录追溯。
- 本次以 PR 记录与 issue 讨论核对为主,未重新执行复现或真机验证。
说明:合入分支与版本以 PR 记录为准;如仍有问题,请重新提 issue 并附版本、shape 与日志。
如果还有后续需要,请重新提 issue。


1 day ago Removed responsible person chenjiayi_
现象
aclnnDeepNorm只校验了meanOut/rstdOut的前导维(x的xDimNum - gammaDimNum维)与维数,没有校验归约维(尾维)必须为 1。把尾维写成非 1 或 0 时,第一段/第二段均返回0/0,kernel 仍按「前导维乘积」个元素写入,导致输出只写一部分、甚至写出声明的元素范围之外。meanOut=[3,1,1](资料要求,合法)[1, 2.2, 3.4],3/3 写入meanOut=[3,1,2](尾维非 1)[1, 2.2, 3.4, <未写>, <未写>, <未写>],只写 3/6meanOut=[3,1,0](零元素)1缺陷组 B 用
mssanitizer --tool=memcheck跑同一 kernel(DeepNorm_03_high_performance_18)报 6 次illegal write of size 4(3 行 × mean/rstd):同族已记录于 #7039(
aclnnDeepNormGrad未校验mean/rstd归约维形状);本 issue 针对前向aclnnDeepNorm,校验函数在norm/deep_norm/op_host/deep_norm_shape_check.h::CheckDeepNormShapeValue()(arch22/arch35 共用),与反向算子不是同一处代码。期望
资料
docs/aclnnDeepNorm.md中meanOut/rstdOut的「使用说明」为:「shape与输入x满足broadcast关系(前几维的维度和输入x前几维的维度相同,前几维表示不需要norm的维度,其余维度大小为1)」;错误码表161002覆盖**「输入和输出的shape不匹配或者不在支持的维度范围内」**。因此尾维不为 1 时应返回161002,不应放行并写入部分或越界数据。环境
npu-smi26.1.1);越界写复现于 Atlas A2 910B2cann/ops-nnf834b617(git fetch后 master),复现日期 2026-10-25复现步骤
repro.cpp,执行g++ -std=c++17 -O2 repro.cpp -o repro -I$ASCEND_HOME_PATH/include -L$ASCEND_HOME_PATH/lib64 -Wl,-rpath,$ASCEND_HOME_PATH/lib64 -lopapi_nn -lopapi -lascendcl -lnnopbase./repro0,且meanOut第 4 个元素起保持金丝雀值-12345mssanitizer --tool=memcheck --log-file=mc.log ./repro复核越界写复现脚本
#include <cstdio> #include <vector> #include "acl/acl.h" #include "aclnnop/aclnn_deep_norm.h" static aclTensor *mk(const std::vector<int64_t> &s, void *dev) { std::vector<int64_t> st(s.size(), 1); for (int64_t i = (int64_t)s.size() - 2; i >= 0; --i) st[i] = s[i + 1] * st[i + 1]; return aclCreateTensor(s.data(), s.size(), ACL_FLOAT, st.data(), 0, ACL_FORMAT_ND, s.data(), s.size(), dev); } static void *up(const std::vector<float> &v) { void *d = nullptr; aclrtMalloc(&d, v.size() * 4, ACL_MEM_MALLOC_HUGE_FIRST); aclrtMemcpy(d, v.size() * 4, v.data(), v.size() * 4, ACL_MEMCPY_HOST_TO_DEVICE); return d; } static void run(const char *tag, std::vector<int64_t> meanShape) { size_t mn = 1, i; for (auto d : meanShape) mn *= (size_t)d; std::vector<float> xd{1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12}, gxd(12, 0.25f); std::vector<float> bd(4, 0.1f), gd(4, 1.5f), yd(12, -12345.0f); std::vector<float> md(mn ? mn : 1, -12345.0f); void *dx = up(xd), *dgx = up(gxd), *db = up(bd), *dg = up(gd), *dy = up(yd); void *dm = up(md), *dr = up(md); std::vector<int64_t> xs{3, 1, 4}, gs{4}; aclTensor *tx = mk(xs, dx), *tgx = mk(xs, dgx), *tb = mk(gs, db), *tg = mk(gs, dg); aclTensor *tm = mk(meanShape, dm), *tr = mk(meanShape, dr), *ty = mk(xs, dy); uint64_t ws = 0; aclOpExecutor *e = nullptr; aclnnStatus r1 = aclnnDeepNormGetWorkspaceSize(tx, tgx, tb, tg, 0.3, 1e-6, tm, tr, ty, &ws, &e); aclError r2 = ACL_SUCCESS; if (r1 == ACL_SUCCESS) { void *w = nullptr; if (ws) aclrtMalloc(&w, ws, ACL_MEM_MALLOC_HUGE_FIRST); r2 = aclnnDeepNorm(w, ws, e, nullptr); aclrtSynchronizeDevice(); } std::vector<float> m(mn ? mn : 1); aclrtMemcpy(m.data(), (mn ? mn : 1) * 4, dm, (mn ? mn : 1) * 4, ACL_MEMCPY_DEVICE_TO_HOST); int keep = 0; for (i = 0; i < mn; ++i) if (m[i] == -12345.0f) keep++; printf("%-24s GetWorkspaceSize=%d Execute=%d meanOut=[", tag, (int)r1, (int)r2); for (float v : m) printf("%.4g ", v); printf("] 未写=%d/%zu\n", keep, mn); } int main() { aclInit(nullptr); aclrtSetDevice(0); run("对照组 [3,1,1] 合法", {3, 1, 1}); run("缺陷组 [3,1,2] 尾维非1", {3, 1, 2}); run("缺陷组 [3,1,0] 零元素", {3, 1, 0}); return 0; }实测输出:
影响
meanOut/rstdOut按归一化维拼错时(例如照搬LayerNorm的[N,1,1]写成[N,1,D]),调用方拿到0返回码,只有前 N 个统计量有效、其余是未初始化内存,而yOut仍按正确语义算出 —— 两份输出互相矛盾且无告警;当声明元素数少于前导维乘积(含[N,1,0])时 kernel 写出声明范围,memcheck 报illegal write。建议
norm/deep_norm/op_host/deep_norm_shape_check.h的CheckDeepNormShapeValue()目前只对i < xDimNum - gammaDimNum的前导维比较rstd/mean与x(第 124-130 行),gammaDimNum段只校验gamma/beta(第 131-137 行),rstd/mean的尾维完全没有参与比较。建议对
i in [xDimNum-gammaDimNum, xDimNum)断言rstd/mean对应维等于 1,不满足则按OP_LOGE_FOR_INVALID_SHAPE_WITH_REASON返回GRAPH_FAILED(映射161002)。该函数被 arch22/arch35 共用,改一处覆盖两侧。