[BUG]libop_abi.so undefined symbol

[yhliu918]

Hi guys, I cloned the latest version of deepmd-kit, master branch and devel branch.
My TF environment is tensorflow-gpu 2.4.1 ( pip source ), with python3.6 , gcc/g++ 7.3.1 , cuda10.1
When I 'pip install .' , it would deliver the error message of "libop_abi.so : undefined symbol: _ZN10tensorflow8OpKernel11TraceStringEPNS_15OpKernelContextEb "

[njzjz]

I cannot reproduce the error on a new environment of Google Colaboratory. https://gist.github.com/b9b7c5c3b002e3e1b6e853c5b0e7317b

[njzjz]

Can you post the output of tensorflow.version.COMPILER_VERSION and tensorflow.version.GIT_VERSION?

[yhliu918]

Hi, my TF version is tensorflow-gpu 2.4.1, I think you didn't use the gpu version in the link you replied

[AnguseZhang]

tensorflow.version.COMPILER_VERSION
'7.3.1 20180303'
tensorflow.version.GIT_VERSION
'v2.4.0-49-g85c8b2a817f'

[njzjz]

It can not be reproduced on tensorflow-gpu as well. Note that tensorflow is also the GPU version.

[njzjz]

I notice tensorflow/text#385 reported a same error, but their tensorflow is from docker source and not pip.

[AnguseZhang]

How to fix this problem : first install tensorflow and then install deepmd-kit.

[hyhuang00]

I tried to solution above but the problem persists:
tensorflow.python.framework.errors_impl.NotFoundError:~/.local/lib/python3.7/site-packages/deepmd/libop_abi.so: undefined symbol: _ZN10tensorflow11GetNodeAttrERKNS_9AttrSliceEN4absl14lts_2020_02_2511string_viewEPSt6vectorIiSaIiEE

[njzjz]

I tried to solution above but the problem persists:

tensorflow.python.framework.errors_impl.NotFoundError:~/.local/lib/python3.7/site-packages/deepmd/libop_abi.so: undefined symbol: _ZN10tensorflow11GetNodeAttrERKNS_9AttrSliceEN4absl14lts_2020_02_2511string_viewEPSt6vectorIiSaIiEE

Please provide more information such as how you install it. If it's a new problem, you can open a new issue.

[njzjz]

In #797, I'm going to throw a message if this error happen.