cjweeks/tensorflow-cmake

Unable to use GPU even though built with cuda option. Any new packages need to be added to BUILD file for tensorflow 1.13 version

deeptf opened this issue · 0 comments

deeptf commented

@cjweeks : This is a great repository. I am still not sure why building tensorflow as a standalone package is super difficult even today. Your repo is of great help.
However I have had two issues:

  1. When I follow your steps and build tensorflow with CUDA options enabled and also --config=monolithic. The inference code in C++ is still not using GPUs. I am unable to list GPU devices also using C++ API.

  2. In versions (master and tf=1.13) there is a problem even for CPU. There are dependencies to external packages (one example: https://github.com/abseil/abseil-cpp). Without these being build, the libtensorflow_all.so complains missing header files. Can you please update any change you think is needed in this regard?

Thanks!