Open
[性能问题][torch._foreach_tanh]torch._foreach_tanh算子存在 性能问题(两者都有) #98
sleeveenAdd on  4 days ago
sleeveenCollaborators
4 days ago Create

确定性问题描述:[性能问题][torch._foreach_tanh]torch._foreach_tanh算子存在 性能问题(两者都有)
通路:torchapi
golden版本:20261103
golden链接:https://gitcode.com/opencann/operator-artifacts/blob/master/cannbot-simt/ForeachTanh/deliverables/thirdparty_golden/golden.py
cann版本:20261103
ttk版本:20261103
用例链接:https://gitcode.com/sleeveen/operator_test_case/tree/main/LongTail/Cases/Functional_Accuracy/torchapi/torch._foreach_tanh
用例名:OP-foreach_tanh-L1-001
复现命令:python3 -m ttk e2e -i 失败用例集_torch._foreach_tanh_torchapi_性能失败.csv --dl=1 --plugin /home/CI/golden/torch._foreach_tanh/ --device-whitelist=0 --xpu-perf -t OP-foreach_tanh-L1-001
问题现象:

均值判断: 均值 = 该算子所有有gpu/npu性能结果的gpu/npu之和(0.453) ÷ 有性能值用例总数(300) = 0.0015; 与用例“带宽折算(均值)”比对, 均值 < 带宽折算(均值)判性能均值不过(本次均值不过用例300条)
单case判断: 逐用例比对 gpu/npu 与“带宽折算(单case)”, gpu/npu < 带宽折算(单case)判性能单case不达标(本次单case不达标用例300条)
均值(自算,共300条有性能用例)=0.0015 < 带宽折算(均值)=1.250, 性能均值不过;gpu/npu=0.005 < 带宽折算(单case)=0.417, 性能单case不达标

【用例 OP-foreach_tanh-L1-001】
2026-11-08 07:13:05,204 [WARNING] [119063 OP-foreach_tanh-L1-001 OnXpuProfiling] [dispatcher.py:705]: 424 retry (1/5): {"error": "missing spec dependency: golden", "missing": "golden"} 

开发责任人:叶钊
PL:叶钊

likedislike