About dependency required installing libtensorflow_cc from conda channel.

[Cloudac7]

Well, it is a dirty trick to just jump the suffering step of compiling TensorFlow C++ interface from installing it from conda. I have used the very way to simplify installation steps for a lot of times, since there might be some protobuf issue using conda to install the whole DeePMD-kit environment for compiling some other interface from it.

However, when I try installation of libtensorflow_cc=2.5.0=gpu_cuda10.1_0 (latest TensorFlow interface version providing support to CUDA 10.1), it raised libiomp5.so is lost without loading Intel MPI environmet as well as Intel MKL, which is not mentioned when using the former compiled 2.3.0 version. It seems also not installed from conda.

So is there any change in dependency of libtensorflow_cc from 2.3.0 or earlier version to 2.5.0, and should it be added for conda? And would it be OK to supply libtensorflow_cc release packages?

By the way, there might be something wrong using libtensorflow_cc=2.5.0 compiling Lammps, raising some namespace error for eigen. More detail error info would add later. Rolling back to 2.3.0 may solve the issue, however for newer Nvidia Ampere Architecture devices, TensorFlow 2.3.0 might cause some issue and I have to upgrade to verison later than 2.4.1 (CUDA>=11.1).

[njzjz]

libomp5 is provided by intel-openmp.