On 2017-12-19 09:16 , Enrico Maria Crisostomo wrote: > Ah ok, I saw this. If I understand it correctly it's a subset of it. > I've also tried to build it from source but I'm not sure at all I would > solve the primary issue: according to the official documentation, the > output of the build is the whl file: > > https://www.tensorflow.org/install/install_sources > > So, even if we update the port to build from source, we end up exactly > with the same issue: how to install the whl file. > > Or perhaps my Python-foo is so weak I'm not seeing something obvious.
This is indeed an unusual situation; most packages on PyPI have a source archive that can be built with 'python setup.py install'. The python portgroup has no support for installing from wheels. As you found, the first complication there is that there is no longer just one distfile… If you have to install a wheel, pip is the tool for doing that. Whether you have to in this case is another question. I haven't tried it, but I wouldn't think their build system would produce *only* a wheel. The files that go into the wheel would probably be there in some build directory too. It looks like building from source has its own complications though. You need Bazel, which needs JDK 8. You also need the CUDA SDK for Nvidia GPU support. - Josh