已开启
[功能问题][AdamApplyOneAssign]AdamApplyOneAssign算子存在 功能问题 #104
sleeveen创建于  1 天前
sleeveen成员
1 天前 创建

确定性问题描述:[功能问题][AdamApplyOneAssign]AdamApplyOneAssign算子存在 功能问题
通路:geir
golden版本:20261113
golden链接:https://gitcode.com/opencann/operator-artifacts/blob/master/AdamApplyOneAssign/Golden/adam_apply_one_assign.py
cann版本:20261113
ttk版本:20261113
用例链接:https://gitcode.com/sleeveen/operator_test_case/tree/main/LongTail/Cases/Functional_Accuracy/GEIR/AdamApplyOneAssign
用例名:AdamApplyOneAssign_L1_tolerance_d_b_014、AdamApplyOneAssign_L1_030
错误码:(未获取到)
复现命令:python3 -m ttk geir -i 失败用例集_AdamApplyOneAssign_geir_功能失败.csv --plugin /home/CI/golden/AdamApplyOneAssign/ --device-whitelist=0 -d -b -c=false --xpu-perf --dl=1 -t AdamApplyOneAssign_L1_tolerance_d_b_014
问题现象:

PROFILE_CRASH | PROFILE_CRASH

CANNBot实际失败分析:
结论(实际日志核查):

  1. 根因类别:算子NPU(GEIR)实现缺陷为主,环境(主机内存不足)为辅。 geir执行后该算子3个输出实际仅返回1个元素(size=1),与golden/三方结果尺寸不符导致比较失败,非golden或用例问题。
  2. 关键证据:
  • ttk.log: [output_parse.py:71] output[0] reshape failed (cannot reshape array of size 1 into shape (21,61,26)); keeping flat array(各用例期望shape不同,输出均为size=1);
  • ttk-compare.log: Output 0 size is different among output/golden/third_party: 1 vs 9989 vs 9989(三个输出均如此);
  • 附带的L1_030日志仅有"run start",被截断(未捕获错误行),与内存告警相关:memory_gate.py:89 Memory low: available 5.9 GB < floor 6.2 GB; waiting up to 300 s、:107 Memory still low after 300 s; proceeding anyway。
  1. 排查建议: ①重点查GEIR图执行侧AdamApplyOneAssign输出shape/描述符传递(疑似更新类in-place输出shape被置为1或仅写1个元素),dump设备输出验证;②检查op实现输出tiling/shape计算;③清理主机内存后复跑L1_030,确认是否叠加OOM导致日志截断。
    开发责任人:叶钊
    PL:叶钊

迁移来源:https://gitcode.com/sleeveen/operator_test_case/issues/260

likedislike
sleeveen成员
1 天前 评论: