daquexian

> oneflow.bernoulli profile测试过程中报错 Error Type: oneflow.ErrorProto.check_failed_error 这个是因为 flow.bernoulli 不支持 cuda tensor，已经汇报给相关同事了，也可以先跳过，我会增强一下 autoprof 的功能支持只测 cpu

> 这个对应python层的函数是什么？ autotest 里的 `random_tensor`

> dry run [Oneflow-Inc/OneTeam#489](https://github.com/Oneflow-Inc/OneTeam/issues/489) > > 现在为什么要去掉了 UndefinedBehaviorSanitizer 发现由于其它地方的变更，dry run 目前处在不能正常工作的状态 https://github.com/Oneflow-Inc/OneTeam/issues/1527#issuecomment-1172134013 😂，和 @jackalcooper 确认之后决定删掉

这个 PR 修复了两个 sanitizer 查出的 bug，其中一个 bug 可能是 ci 里有概率出现却一直没找到原因的 double free 现象的原因

I confirmed it crashed on iPhone 6. And after trying some times it ran successfully but the result it got is never correct.

为了稳妥起见，加了一个 ONEFLOW_PYTHON_STACK_GETTER 环境变量，需要用户显式打开

在 https://github.com/Oneflow-Inc/oneflow/pull/9534 中继续开展

Thanks for your report! The issue you reported does exist. For most models, the redundant constant ops and initializers will be eliminated in the following procedure -- onnx optimizer. But...

> Thanks for following up @daquexian ! Indeed, my current fix does not care about subgraphs, but I believe it could be easily modified to do so. When calculating `used_tensors...

> After I converted my pth(96,544kb) to .onnx(98,446kb) , the error occured and after simplifying,the new onnx(97,607 kb) can't work as well and the error still exsited. Sorry for the...