Bug 922374 – Last-rite tensorflow associated ecosystem

• Product: Gentoo Linux
• Status: CONFIRMED ---

[Description David Seifert 2024-01-18 12:05:30 UTC]

TensorFlow in Gentoo is practically unmaintained and should be last-rited at this point. The number of bugs is orders of magnitude more than the packages using it, and the maintainer has made 4 commits last year. The following packages should be last-rited:

dev-build/bazel
sci-libs/keras
sci-libs/tensorflow
sci-libs/tensorflow-estimator
sci-visualization/tensorboard

[Comment 1 Larry the Git Cow 2024-01-18 12:53:37 UTC]

The bug has been referenced in the following commit(s):

https://gitweb.gentoo.org/repo/gentoo.git/commit/?id=8968f9287c683d0fc9d46b87e4ac624c7c7b066e

commit 8968f9287c683d0fc9d46b87e4ac624c7c7b066e
Author:     David Seifert <[e-mail]>
AuthorDate: 2024-01-18 12:52:21 +0000
Commit:     David Seifert <[e-mail]>
CommitDate: 2024-01-18 12:52:21 +0000

    package.mask: Last rite dev-build/bazel, sci-libs/keras, sci-libs/tensorflow, sci-libs/tensorflow-estimator, sci-visualization/tensorboard
    
    Closes: https://github.com/gentoo/gentoo/pull/34881
    Bug: https://bugs.gentoo.org/922374
    Signed-off-by: David Seifert <[e-mail]>

 profiles/package.mask | 19 +++++++++++++++++++
 1 file changed, 19 insertions(+)

[Comment 2 Dennis Jenkins 2024-01-19 23:29:16 UTC]

I use tensorflow on Gentoo, specifically form a custom C++ app.  The python install is not useful to me.  I will be quite sad if tensorflow is removed from my favorite Linux distro (been using Gentoo since 2004).

[Comment 3 Sam James 2024-01-19 23:34:19 UTC]

(In reply to Dennis Jenkins from comment #2)
> I use tensorflow on Gentoo, specifically form a custom C++ app.  The python
> install is not useful to me.  I will be quite sad if tensorflow is removed
> from my favorite Linux distro (been using Gentoo since 2004).

Yes, I can understand that. But it also doesn't work right now, doesn't build, and see the huge list of bugs. Simply not removing it doesn't make those go away.

[Comment 4 David Seifert 2024-01-20 11:45:29 UTC]

My suggestion is to move this to an overlay (say ::tensorflow), since these 5 packages are well-contained and don't have any major "depgraph sprawl" like other overlays that tend to take over your system. Adding that overlay would be easy and there wouldn't be any knock-on effects.

[Comment 5 David Seifert 2024-01-20 11:46:55 UTC]

For reference, Debian's packaging situation and experience mirrors a lot of the pains we've been having too: https://wiki.debian.org/MachineLearning/Tensorflow

[Comment 6 Nuno Alexandre 2024-01-23 16:16:16 UTC]

I'm a long time gentoo user too and I too need this.
I understand it is not easy to maintain this package, but it feels like a big step back in AI for gentoo if this is dropped. If these packages start being shipped in an overlay, that's fine. We can then tinker and get it working on our machines. Eventually, I'm sure a way to build this consistently will be found. And by then we can move it to the oficial gentoo repo.
Btw, the 2.12.0 version works on my machine, built with the 5.3.0 bazel (built in 16/11/2023) with gcc:11, cuda-toolkit 11.8.0 and cudnn 8.6.0.163. In my experience tensorflow is very sensitive to what versions and compilers you use, so watch out for that. Also, make sure dependencies are also compiled with the same compiler you are using for tensorflow.

Removing it outright is simply saying we're never going to support this and couldn't be bothered with it.

[Comment 7 Nuno Alexandre 2024-01-23 18:50:53 UTC]

I just recompiled it successfully.
Compiled with gcc:11, bazel 5.3.0, cuda-toolkit 11.8.0 and cudnn 8.6.0.163. Bazel needs dev-java/openjdk:11 or equivalent, make sure that is set as well.
Make sure that related packages are also compiled with gcc:11 and have valid ABI versions.

[Comment 8 Eli Schwartz 2024-01-24 02:56:27 UTC]

(In reply to Nuno Alexandre from comment #6)
> In my experience tensorflow is very sensitive to [...]
> so watch out for that.

Yes, I believe this was part of the original mask message... are you speaking in favor of keeping it, or of removing it?

Because I think the point of the Devs proposing to drop it was *precisely* that they consider having to "watch out for that" to be an unconscionable drain upon Gentoo, 

> Removing it outright is simply saying we're never going to support this and
> couldn't be bothered with it.

Absolutely not. Failing to ever package it could be taken to mean "we couldn't be bothered with it". Trying, packaging it once, and then removing it, says "we tried, we really could be bothered, but in the end we lost the struggle".