Open
[功能问题][ForeachAbs]ForeachAbs算子存在 功能问题 #100
sleeveenAdd on  3 days ago
sleeveenCollaborators
3 days ago Create

确定性问题描述:[功能问题][ForeachAbs]ForeachAbs算子存在 功能问题
通路:geir
golden版本:20261109
golden链接:https://gitcode.com/opencann/operator-artifacts/blob/master/cannbot-simt/ForeachAbs/deliverables/thirdparty_golden/golden.py
cann版本:20261109
ttk版本:20261109
用例链接:https://gitcode.com/sleeveen/operator_test_case/tree/main/LongTail/Cases/Functional_Accuracy/GEIR/ForeachAbs
用例名:ForeachAbs_L1_003、ForeachAbs_L1_013
错误码:(未获取到)
复现命令:python3 -m ttk geir -i 失败用例集_ForeachAbs_geir_功能失败.csv --device-whitelist=0 -d -b -c=false --golden-mode Disable -t ForeachAbs_L1_003
问题现象:

【用例 ForeachAbs_L1_003】
2026-11-11 19:24:03,552 [ERROR] [438484 ForeachAbs_L1_003 OnGenerateGolden] [invoke.py:381]: Golden generation failure 
+-------------------------------------------+
| MODE     | Performance  | Comparison      |
+-------------------------------------------+
| CST_BIN  | CST_OFF      | CST_OFF         |
+-------------------------------------------+
| DYN_BIN  | 73.196       | GOLDEN_FAILURE  |
+-------------------------------------------+
| STATUS   | FAIL         | FAIL            |
+-------------------------------------------+ 

CANNBot实际失败分析:
1) 失败根因与错误类别:golden问题(golden脚本实现缺陷),非NPU算子缺陷。
Golden生成阶段即抛异常(GOLDEN_FAILURE),失败发生在CPU侧golden计算前的数据转换,NPU侧DYN_BIN已正常执行(perf 73.196us,MD5三次一致),与算子NPU实现无关。所有失败用例(L1_003等22个)输入dtype均为bfloat16,属同一根因。
2) 关键日志证据:

ttk.log:7  [ERROR] Golden generation failure
ttk.log:18-21  File "/home/CI/golden/ForeachAbs/golden.py", line 48, in golden
  x_list = [torch.from_numpy(np.ascontiguousarray(t)) for t in x]
  TypeError: can't convert np.ndarray of type ml_dtypes.bfloat16.
  The only supported types are: float64, float32, float16, ... and bool.
ttk.log:32  DYN_BIN | 73.196 | GOLDEN_FAILURE
ttk-compare.log:5  Output 0 Precision is GOLDEN_FAILURE (golden sentinel)

3) 初步定位建议:

  • golden.py:48 直接用 torch.from_numpy(np.ascontiguousarray(t)) 处理ml_dtypes.bfloat16数组,torch不支持该numpy dtype,需修复转换逻辑;
  • 建议改法:先用 t.view(uint16) 提取原始位再转torch bfloat16(框架内若有bfloat16辅助转换接口则优先复用),或统一转float32再由框架按bfloat16精度比对;
  • 检查测试框架输入生成是否统一使用ml_dtypes.bfloat16,确认同类foreach系列算子golden是否均存在此隐患,一并修复。
    开发责任人:叶钊
    PL:叶钊
likedislike
sleeveenCollaborators
3 days ago Comments:
SsleeveenCollaborators
2 days ago Added label: 已提单