Official release of tensorflow==1.15.2 does not include GPU support

[pavanky]

System information

• OS Platform and Distribution: Linux
• Mobile device (e.g. iPhone 8, Pixel 2, Samsung Galaxy) if the issue happens on mobile device:
• TensorFlow installed from (source or binary): binary
• TensorFlow version: 1.15.2
• Python version: python2, python3.6
• Installed using virtualenv? pip? conda?: pip

Describe the problem

As announced Tensorflow 1.15 contained GPU support by default. Tensorflow 1.15.2 no longer has GPU support.

[pavanky]

Related:

• There is no 1.15.2 version of tensorflow_cpu: https://pypi.org/project/tensorflow-cpu/#history
• The Java version of 1.15.2 hasnt been released either: https://mvnrepository.com/artifact/org.tensorflow/libtensorflow_jni

[mihaimaruseac]

There has been an unfortunate change in our build scripts and 1.15.2 uses tensorflow_gpu for GPU and tensorflow for CPU.

[pavanky]

@mihaimaruseac This isnt a blocker, but it would nice if you can make an announcement about this at least.

[amahendrakar]

@pavanky,
Is this still an issue? Please please free to close it if if answers your query.

[pavanky]

@amahendrakar I can build the combined wheel if I build from source, but this is about the official release.

I dont see the download page updated yet.

[mihaimaruseac]

For 1.15 what our scripts did was:

1. produce tensorflow_gpu and tensorflow pip package
2. upload tensorflow_gpu pip package as it is
3. rename tensorflow pip package to tensorflow_cpu and upload it under tensorflow_cpu
4. rename tensorflow_gpu as tensorflow and upload it under tensorflow

Both of these renames are more than just a cp source.whl target.whl, they also required changing the metadata in the wheel. But this process caused issues with integrity checks as filenames changed.

After 1.15.0 was released (actually, just before 2.1.0 was released) we fixed all of those issues but the fix was not backported into r1.15.

Currently, one could think we could attempt to manually do the steps above. Except we cannot. Pypi doesn't allow us to reuse package names (for good reasons) so once we have uploaded tensorflow_gpu==1.15.2 and tensorflow==1.15.2 we cannot change them around.

Actually, it is better to not do this since the move above caused tensorflow_estimator to depend on tensorflow_cpu_estimator for example (due to a bug). So if we wanted to manually do the steps above then we would need to create 3 more patch releases with no actual changes (2 for tensorflow, 1 for estimator).

We will update documentation on download page though

[pavanky]

@mihaimaruseac updating the documentation sounds good. this can be closed at that point.

[mihaimaruseac]

Documentation updated. Please let me know if there's something else to change.

Apologies again for not having a single pip anymore on 1.15

[wookayin]

With all due respect I think we should release a hotfix ASAP (say 1.15.3, 1.15.2.1 or something) to provide GPU support. Also, we could remove the 1.15.2 binary from pypi (as we did with 1.15.1). I believe it is obviously an unintended behavior and therefore a bug. I don't think it makes a sense to have a very unconsistent package scheme in the same version series. UPDATED: Please see the comments below.

Due to the promise of v1.15, some people (definitely in certain cases) might have a dependency specification like tensorflow>=1.15 rather than tensorflow-gpu>=1.15 in the hope that GPU is enabled. But, in such cases we might end up installing a package without GPU (which I did).

[wookayin]

I carefully went through the documentation and the history again, and now I think it also makes a sense to revert the package consolidation behavior for 1.15.x series (which also makes a sense as this is kinda breaking change around v2.0). Given that the documentation has been already updated, we should keep this one.

Only remaining issues would be legacy package specification in some cases, or outdated announcements.

UPDATE 1: https://www.tensorflow.org/install/gpu still says 1.15 has a consolidated package (which needs to be updated).

Thanks!

[metaxmx]

Hi, I also got confused and used the wrong package (tensorflow==1.15.2 without GPU support), because I read this section in the install documentation:
https://www.tensorflow.org/install/gpu#older_versions_of_tensorflow

For the 1.15 release, CPU and GPU support are included in a single package:
pip install --pre "tensorflow==1.15.*"

Please add a remark to this section, that it does not apply to the version 1.15.2 (and upwards). It's quite confusing if you explicitly state that the 1.15.x versions do NOT use a separate "-gpu" package, when in fact some of them do.

[mihaimaruseac]

Should be handled soon (with a 24hours delay due to the upgrade process). Apologies for missing that page

[jingweimo]

@mihaimaruseac Which cuda version is compatible with tensorflow-gpu==1.15.2 in windows?

[mihaimaruseac]

It should be the same CUDA version as 2.0.

[jingweimo]

@mihaimaruseac I have tested that it works with cuda 10.0