@mrry @asimshankar can comment. I think it's safer to assume that it won't be there in the next few months.

We don't build GPU-enabled binaries for the C or JNI libraries for Windows as part of the release process yet.
However, it should happen within a release or two (probably 1.7).

We just started setting up the nightly build for GPU - https://ci.tensorflow.org/view/Nightly/job/nightly-libtensorflow-windows-gpu/
(which hasn't had a successful build yet :)

Ok thanks, I'm glad it's in your plans :)

Do you have an idea how long it will be until 1.7?

Thanks,

Nagging Assignee: It has been 14 days with no activity and this issue has an assignee. Please update the label and/or status accordingly.

Nagging Assignee @asimshankar: It has been 14 days with no activity and this issue has an assignee. Please update the label and/or status accordingly.

Hi,
You mentioned that there should be windows gpu java support with 1.7, I see the 1.7 release, is there support for gpu in java with windows coming soon?

@jedakiah13 : Unfortunately, I was not able to get this in for 1.7.0-rc0 .

so are we likely to have to wait until 1.8?

I was curious to know the same, and appreciate the updates on the status of this @asimshankar !

I too was wondering what the status on this is.

Unfortunately, nope, we didn't get this in time for 1.7.
However @case540 has been looking into the build issues, so hopefully we'll be able to make progress in the next release or two.

Nagging Assignee @case540: It has been 14 days with no activity and this issue has an assignee. Please update the label and/or status accordingly.

Nagging Assignee @case540: It has been 29 days with no activity and this issue has an assignee. Please update the label and/or status accordingly.

Any news here? ETA?

Still waiting for an update... We would love to have this up and running.

Windows libtensorflow GPU build is broken and requires a fix from the protobuf team. Not entirely sure on eta but it is something we are working on.

@case540 What exactly needs to be done?

Nagging Assignee @case540: It has been 15 days with no activity and this issue has an assignee. Please update the label and/or status accordingly.

Nagging Assignee @case540: It has been 32 days with no activity and this issue has an assignee. Please update the label and/or status accordingly.

Still waiting for an update....

Same here... @case540 are you working on this?

I'm working on this. It should be available soon.

@jedakiah13 @tlf30 - could you try with 1.10.0?

@av8ramit thanks a lot for your work!
I can see the windows-x86_64\tensorflow_jni.dll in libtensorflow_jni-1.10.0-rc1.jar but not in libtensorflow_jni_gpu-1.10.0-rc1.jar.
Should the lib jni_gpu include a native win lib as well?
Currently gpu for win does not work for me referencing both libs via Maven.

@mstritt - Please try version 1.10.0, not 1.10.0-rc1 :)

@asimshankar , thanks for the hint, indeet it's included there.
Unfortunately I get a dependency missing error now (see log).
msvcp140.dll is available, but I have cudbb64_6.dll. Do I have to upgrade to cudnn 7 ?

org.tensorflow.NativeLibrary: tryLoadLibraryFailed: no tensorflow_jni in java.library.path
org.tensorflow.NativeLibrary: jniResourceName: org/tensorflow/native/windows-x86_64/tensorflow_jni.dll
org.tensorflow.NativeLibrary: frameworkResourceName: org/tensorflow/native/windows-x86_64/tensorflow_framework.dll
org.tensorflow.NativeLibrary: org/tensorflow/native/windows-x86_64/tensorflow_framework.dll not found. This is fine assuming org/tensorflow/native/windows-x86_64/tensorflow_jni.dll is not built to depend on it.
org.tensorflow.NativeLibrary: extracting native library to: C:\Users\[...]\AppData\Local\Temp\tensorflow_native_libraries-1534235673406-0\tensorflow_jni.dll
org.tensorflow.NativeLibrary: copied 158330368 bytes to C:\Users\[...]\AppData\Local\Temp\tensorflow_native_libraries-1534235673406-0\tensorflow_jni.dll
Exception in thread "main" java.lang.UnsatisfiedLinkError: C:\Users\[...]\AppData\Local\Temp\tensorflow_native_libraries-1534235673406-0\tensorflow_jni.dll: Can't find dependent libraries

I think so, see https://www.tensorflow.org/install/install_windows#requirements_to_run_tensorflow_with_gpu_support

ok, works like a charm with cuda9+cudnn7. Thanks a lot for this!

No problem @mstritt !

@asimshankar my team and I will check it out, thank you.

@mstritt Could you please tell the version of tensorflow too?

@carlosuc3m : it is working since TF 1.10

@mstritt it is not working for me in tf 1.12, with cuda 9 and cudnn 7.
It says that java.lang.UnsatisfiedLinkError: tensorflow_jni.dll: Can't find dependent libraries.
Thanks for your answer

I have also tried with TF 1.15 and it loads well but it seems that is still using the CPU, how can I check this? Or should I call any different method? Regards

Just in case someone finds it useful, the problem that was causing this error was that I was not using the correct CUDA version. I tried for tf 1.12 and tf 1.15. For tf 1.12 I needed CUDA 9.0 and for tf 1.15 I needed CUDA 10.0

Hi all.
My config is Win 10 + CUDA 10.1 + cuDNN 7.6 + (may this have an interest, Python 3.7.7 + tensorflow 2.1 installed with pip in a Python virtual env; TF works well in Python).
In Java, which I hope having no dependency with TF in Python, I use without problem TF 1.4 and TF 1.14 for cpu with ant. I simply put tensorflow_jni.dll in the project subdirectory myjni and mention -Djava.library.path=myjni and it runs smoothly.
When with ant, I use the same java sample program with TF 1.14 for gpu, with the same -Djava.library.path=myjni and tensorflow_jni.dll (gpu version, bigger) in the myjni directory, I receive the following error:
Exception in thread "main" java.lang.UnsatisfiedLinkError: C:\Users\<me>\AppData\Local\Temp\tensorflow_native_libraries-1584890522116-0\tensorflow_jni.dll: Can't find dependent libraries
What is strange is that each run generates a new directory C:\Users\<me>\AppData\Local\Temp\tensorflow_native_libraries-<NUMBER>\ where <NUMBER> is each time different. The directory is removed automatically after about 2 sec (for the rare case it run fine, the directory is not removed). I would like to understand why.
Hence, tensorflow_jni.dll in myjni directory is ignored.
So many days spent here. Any clue?
NB: I also have:

• (8!) Microsoft Visual C++ 20nn Redistribuable (x64,x86) with nn in [05, 12, 13, 15-19]. I don't know if I should remove some of them...
• Microsoft Visual Studio Community 2015 with Updates