Lack of maintenance

[hajapy]

This feedstock has stagnated relative to the upstream project, which I believe is largely due to how complex tensorflow is to build within conda. We have also failed to really solve the gpu/cuda case which is most critical for this project.

I’m wondering what next steps people think we should take.

This is a highly downloaded package - it would be great for the community if we could keep it up to date and deliver on variants with gpu support.

[isuruf]

@derekschinke, complaining doesn't help. PRs are always welcome.

[hajapy]

This issue is intended to be a call for help and input. We have had a few fantastic contributors advance the state of this feedstock, but large challenges remain unsolved.

Defaults does seem to have a better solution, but that seems like a good chunk of work for @jjhelmus and others to maintain. Perhaps we could take advantage of some of their solution here?

Maybe we could engage with tensorflow devs and/or nvidia more actively to see if there is any support they could lend?

[adrianchia]

I was trying to rebuild the tensorflow 2.x based on https://github.com/AnacondaRecipes/tensorflow_recipes, a few things I noticed is that:

1. it uses Bazel 3.1.0, which is available in the default conda channel and available for Linux and Windows, but not OSX

2. build time could take more than 6 hours, which is longer than the maximum provided by azure pipelines.

[njzjz]

A few days ago, I successfully built both CPU and GPU version of TensorFlow 2.3 C++ interface, i.e. libtensorflow_cc, based on the default channel. I upload my recipe to https://github.com/deepmd-kit-recipes/libtensorflow_cc-feedstock/tree/master/recipe. It may be useful if anyone is interested in it. I created the repository with conda-forge build tools, but I finally built it on my computer since the build time is too long.

[asford]

@hajapy Would you all be open to more limited maintenance to help cover the cases that aren't being tackled by the current defaults? For example, pushing the existing OSX packaging strategy here ("repackage the upstream wheel build") forward for the 2.* release series?

Recreating the hard work already present in defaults for the windows and linux builds seems to be of limited value, but there's a big gap in tensorflow>=2-on-osx-64-on-conda right now we could tackle.

[isuruf]

@asford, IMO, that's fine. Please send a PR even if it's not working fully.

[xhochy]

Fixed by #110