Closed
[Bug-Report|缺陷反馈]: aclnnInplaceAddRmsNorm 空 tensor 未被拦截,两段均返回 0(资料明写不支持空 Tensor) #7398
liuhanwen8Add on 26 days agoClosed at 6 days ago
26 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:文档反馈;aclnnInplaceAddRmsNorm 空 tensor 未被拦截,两段均返回 0(资料明写不支持空 Tensor)。
当前判断:从描述看,这是资料/注释与实现口径不一致类文档反馈「aclnnInplaceAddRmsNorm 空 tensor 未被拦截,两段均返回 0(资料明写不支持空 Tensor)」。需要对照接口文档、头文件注释与实现/实测结果核对哪一侧需修正。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:aclnnInplaceAddRmsNorm、aclnnInplaceAddRmsNormGetWorkspaceSize、aclnnAddRmsNorm
下一步:我们会先核对文档声明与实现/测试结果是否一致,再决定改资料还是改代码。


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


25 days ago Assign liuyun_nj as responsible person
25 days ago Removed responsible person liujie12345678
11 days ago Assign liujie12345678 as responsible person, remove original responsible person: liuyun_nj
11 days ago Assign liujie12345678 as responsible person, remove original responsible person: liuyun_nj
10 days ago Assign liuyun_nj as responsible person
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:
该 issue 已关闭,关闭记录中暂无关联修复 PR 或结论说明。
处理结果:
- 关闭动作未附带修复摘要、版本说明或验证结论。
验证:
- 基于 issue 讨论与关闭记录核对;本次未重新执行复现。
说明:如需继续跟进,请重新提 issue,并附版本、shape 与复现日志。
如果还有后续需要,请重新提 issue。


1 day ago Removed responsible person chenjiayi_
Describe the current behavior / 问题描述 (Mandatory / 必填)
接口资料在参数表里对
x1Ref、x2Ref、gamma、rstdOut四个张量都写了「不支持空 Tensor」,但实测把归一化维压成 0(x1Ref=[2,0]、gamma=[0]、rstdOut=[2,0])时:aclnnInplaceAddRmsNormGetWorkspaceSize返回0aclnnInplaceAddRmsNorm返回0没有任何拒绝。也就是说「不支持空 Tensor」这条契约在实现里没有落地:调用方带着空张量调用会得到「成功」,而不是文档承诺的错误。
同族接口
aclnnAddRmsNorm的 tiling 里对空 tensor 有显式判断(x1 cannot be an empty tensor),本接口的空张量输入未走到该拒绝路径。Environment Information / 环境信息 (Mandatory / 必填)
Steps to reproduce the issue / 重现步骤 (Mandatory / 必填)
repro_empty.cpp(只用 CANN 公开头文件与libascendcl/libopapi/libnnopbase)。g++ -std=c++17 -O2 -I$ASCEND_HOME/include -I$ASCEND_HOME/include/aclnnop repro_empty.cpp -L$ASCEND_HOME/lib64 -lascendcl -lopapi -lnnopbase -Wl,--allow-shlib-undefined -o repro_empty。source $ASCEND_HOME/set_env.sh后执行./repro_empty。[2,16]/[16]/[2,1]),缺陷组空张量([2,0]/[0]/[2,0]),其余参数相同(fp32、epsilon=1e-6)。#include <cstdio> #include <vector> #include "acl/acl.h" #include "aclnnop/aclnn_inplace_add_rms_norm.h" static int64_t Numel(const std::vector<int64_t>& s) { int64_t n = 1; for (auto v : s) { n *= v; } return n; } static aclTensor* Make(const std::vector<int64_t>& shape, float v, void** p) { int64_t n = Numel(shape); size_t bytes = (size_t)(n > 0 ? n : 1) * sizeof(float); if (aclrtMalloc(p, bytes, ACL_MEM_MALLOC_HUGE_FIRST) != ACL_SUCCESS) { return nullptr; } std::vector<float> h((size_t)(n > 0 ? n : 1), v); aclrtMemcpy(*p, bytes, h.data(), bytes, ACL_MEMCPY_HOST_TO_DEVICE); std::vector<int64_t> dims = shape.empty() ? std::vector<int64_t>{1} : shape; std::vector<int64_t> st(dims.size(), 1); for (int64_t i = (int64_t)dims.size() - 2; i >= 0; i--) { st[(size_t)i] = dims[(size_t)i + 1] * st[(size_t)i + 1]; } return aclCreateTensor(dims.data(), dims.size(), ACL_FLOAT, st.data(), 0, ACL_FORMAT_ND, dims.data(), dims.size(), *p); } static void Case(const char* tag, std::vector<int64_t> xs, std::vector<int64_t> gs, std::vector<int64_t> rs) { aclrtStream stream = nullptr; aclrtCreateStream(&stream); void *p1 = nullptr, *p2 = nullptr, *pg = nullptr, *pr = nullptr; aclTensor* x1 = Make(xs, 1.0f, &p1); aclTensor* x2 = Make(xs, 2.0f, &p2); aclTensor* g = Make(gs, 0.5f, &pg); aclTensor* r = Make(rs, 0.0f, &pr); uint64_t ws = 0; aclOpExecutor* ex = nullptr; int ret = (int)aclnnInplaceAddRmsNormGetWorkspaceSize(x1, x2, g, 1e-6, r, &ws, &ex); printf("[%s] GetWorkspaceSize = %d\n", tag, ret); if (ret == 0) { void* wa = nullptr; if (ws > 0) { aclrtMalloc(&wa, ws, ACL_MEM_MALLOC_HUGE_FIRST); } printf("[%s] Execute = %d\n", tag, (int)aclnnInplaceAddRmsNorm(wa, ws, ex, stream)); aclrtSynchronizeStream(stream); if (wa) { aclrtFree(wa); } } aclDestroyTensor(x1); aclDestroyTensor(x2); aclDestroyTensor(g); aclDestroyTensor(r); aclrtFree(p1); aclrtFree(p2); aclrtFree(pg); aclrtFree(pr); aclrtDestroyStream(stream); } int main() { aclInit(nullptr); aclrtSetDevice(0); Case("对照组 CTRL", {2, 16}, {16}, {2, 1}); Case("缺陷组 DEFECT", {2, 0}, {0}, {2, 0}); aclrtResetDevice(0); aclFinalize(); return 0; }Describe the expected behavior / 预期结果 (Mandatory / 必填)
按资料「不支持空 Tensor」,空张量输入应当在第一段接口被拒绝(返回
161002一类参数校验错误),与同族aclnnAddRmsNorm对空 tensor 的处理保持一致。Related log / screenshot / 日志 / 截图 (Mandatory / 必填)
实测输出(Ascend 910B3 / CANN 9.2.0):
补充观测:
x1Ref=fp32[2x0]、gamma=fp32[0]、rstdOut=fp32[2x0])结果一致:两段均返回 0;x1Ref=[0,16](前导维为 0)时会被拒绝,说明只有「归一化维为 0」这一形态漏了拦截;x1Ref/x2Ref/rstdOut元素数均为 0,没有非空输出被写坏,因此是契约漂移而非脏数据。Special notes for this issue/备注 (Optional / 选填)
aclnnAddRmsNorm的空 tensor 判断对齐。