Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Build tensorflow from source not whl files. #6

Closed
jjhelmus opened this issue May 20, 2016 · 11 comments
Closed

Build tensorflow from source not whl files. #6

jjhelmus opened this issue May 20, 2016 · 11 comments

Comments

@jjhelmus
Copy link
Collaborator

Currently the tensorflow recipe is repackaging the whl files provided by Google into conda packages. The whl files link against a newer version of glibc which limited what Linux distribution that the conda package works on and are only available for Python 2.7 and 3.4 on Linux.

A "better" solution would be to build the package from source, but this requires installing Bazel which requires Java JDK 8. I have successfully built a conda package for 0.8.0 from source using a Ubunutu 14.04 Docker container with Java+Bazel. We could do the same here but it would require quite a bit of customization of this feedstock. Anyone have opinion on this topic?

@rgommers
Copy link

There's 3 people reporting at numpy/numpy#8076 that installing tensorflow breaks numpy, and it looks related to this wheel repackaging business.

@rgommers
Copy link

This comment says that indeed tensorflow pulls in a numpy wheel somehow. So that's pretty badly broken.

@jakirkham
Copy link
Member

Ouch! Yeah the build situation here is not great. It is unfortunately quite difficult to build tensorflow from source. That said, @jjhelmus added bazel recently (merged today), which should help.

@hajapy
Copy link
Contributor

hajapy commented Jul 14, 2017

I'm a little confused, our build recipe installs tensorflow wheel via pip with --no-deps, so the only numpy should be one specified by the meta.yaml run time requirements.

@rgommers
Copy link

I also don't really understand it, but it's reported by 4 people so should be reproducible.

@jakirkham
Copy link
Member

Of course, I think it is more surprise at the issue's existence than any doubt that people are running into an issue. Though you make a good point, steps to reproduce the issue would be very helpful in debugging it. Was that in the issue and I simply missed it?

@rgommers
Copy link

No, they all said "installed tensorflow with conda, then numpy broke". Asked for exact install commands.

@jakirkham
Copy link
Member

Great, thanks Ralf.

@pgunn
Copy link

pgunn commented May 1, 2019

Is it possible to copy the build instructions from the non-conda-forge builds of tensorflow? Those versions don't seem to have the GLIBC version dependency (tested on RHEL7, which has a very dated glibc).

@jakirkham
Copy link
Member

There is some work for this in PR ( #69 ). That said, this is not easy work. So please feel free to chip in if it is of interest and you have cycles.

sodre added a commit to sodre/tensorflow-probability-feedstock that referenced this issue Jun 1, 2019
@jjhelmus
Copy link
Collaborator Author

jjhelmus commented Aug 8, 2019

Done in #74

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

5 participants