Open
[性能问题][INTrainingReduceV2]INTrainingReduceV2算子存在 性能问题(单case失败) #206
sleeveenAdd on  16 days ago
sleeveen
sleeveenCollaborators
16 days ago Create

确定性问题描述:[性能问题][INTrainingReduceV2]INTrainingReduceV2算子存在 性能问题(单case失败)
通路:geir
golden版本:20261028
golden链接:https://gitcode.com/opencann/operator-artifacts/blob/master/hpc-group/july/InTrainingReduceV2/04_golden/golden.py
cann版本:20261028
ttk版本:20261028
用例链接:https://gitcode.com/sleeveen/operator_test_case/tree/main/LongTail/Cases/Functional_Accuracy/GEIR/INTrainingReduceV2
用例名:InTrainingReduceV2_L1_accuracy_perf_d_b_005
复现命令:python3 -m ttk geir -i 失败用例集_INTrainingReduceV2_geir_性能失败.csv --plugin /home/CI/golden/INTrainingReduceV2/ --device-whitelist=0 -d -b -c=false --xpu-perf --dl=1 -t InTrainingReduceV2_L1_accuracy_perf_d_b_005
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(190.415) ÷ 有性能值用例总数(300) = 0.6347; 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0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88条)
gpu/npu=0.037 < 带宽折算(单case)=0.185, 性能单case不达标

FAIL | {'torch': {'status': 'PASS', 'api': 'INTrainingReduceV2KernelSpec', 'device_us': 52.4, 'peak_memory_mb': 62.938112}}

开发责任人:(未获取到)
PL:张磊

likedislike
raoliang_sacraoliang_sacCollaborators
16 days ago Associated pull request
raoliang_sac
raoliang_sacCollaborators
13 days ago Comments:

已确认 性能有问题

likedislike
raoliang_sacraoliang_sacCollaborators
12 days ago Associated pull request