tensorflow-windows-wheel icon indicating copy to clipboard operation
tensorflow-windows-wheel copied to clipboard

Tensorflow 1.8 with CUDA 9.2

Open XuYunqiu opened this issue 6 years ago • 22 comments

Have you compiled Tensorflow 1.8 with CUDA 9.2?

XuYunqiu avatar May 18 '18 08:05 XuYunqiu

I will use CUDA 9.2 to build tensorflow-gpu 1.9.0.

fo40225 avatar May 18 '18 13:05 fo40225

yes and the last cudnn drivers are Download cuDNN v7.1.4 (May 16, 2018), for CUDA 9.2. Please do it with those. Because I cannot find a download link cuDNN v7.1.3

also with avx2 support would be great :)

Thanks for your work!

elranu avatar May 20 '18 01:05 elranu

cuDNN 7.1.3 for CUDA 9.1 Windows 10

cudnn-9.1-windows10-x64-v7.1.zip

SHA1: 914006C6407256408EE9C18D6115FDA1209C2F0B

Google Drive

百度盤

fo40225 avatar May 20 '18 07:05 fo40225

I have rebuild the tensorflow-gpu 1.8.0 with CUDA 9.2& cuDNN 7.1.4.

https://github.com/fo40225/tensorflow-windows-wheel/tree/master/1.8.0/py36/GPU/cuda92cudnn71sse2

try it.

fo40225 avatar May 20 '18 07:05 fo40225

@fo40225 thanks for the last version of cuda 9.2. It works to me much better that the 9.1.

Can you compile me the version with avx support? I think it will work even better ;)

Thanks for your work! I really appreciate it :)

elranu avatar May 23 '18 14:05 elranu

Tensorflow team has planed to use CUDA 9.2 to build 1.9.0, it will contain AVX support.

fo40225 avatar May 23 '18 14:05 fo40225

Offtopic: CUDA 9.1 + AVX (not AVX2) to build 1.8.0 maybe?

LSinev avatar May 23 '18 15:05 LSinev

Nvidia didn't provide cuDNN 7.1.x for CUDA 9.1, maybe cuda 9.1 is deprecated.

fo40225 avatar May 23 '18 19:05 fo40225

Have you compiled Tensorflow 1.8 with CUDA 9.2 in Linux for example ubuntu?

gaohongfein avatar May 28 '18 20:05 gaohongfein

@gaohongfein No, you can ask others who maintains custom Linux build for help.

https://github.com/mind/wheels https://github.com/lakshayg/tensorflow-build https://github.com/yaroslavvb/tensorflow-community-wheels/issues

fo40225 avatar May 29 '18 05:05 fo40225

Thank you very much

gaohongfein avatar May 29 '18 06:05 gaohongfein

@gaohongfein I build the cuda9.2 for TensorFlow1.8.0(python3.6)on ubuntu18.04。 https://github.com/mtianyan/tensorflow-linux-wheel

mtianyan avatar Jun 06 '18 06:06 mtianyan

@fo40225 Ask something about cuda92cudnn71sse2 wheel. Is there AVX2 support on it? // And thanks a lot for compiling these packages, later i would try c++ lib.

sleeplessai avatar Jun 07 '18 02:06 sleeplessai

CMake Error at CMakeLists.txt:322 (message): Selected compiler (or version) is not supported for CUDA I use win10+vs2017+cuda9.2+cudnn7.1 compile tensorflow1.8. When I use cmake generate the file, it get the error.How can I resolve?

yanhaiming56 avatar Jul 02 '18 01:07 yanhaiming56

@yanhaiming56 You have two options to solve that:

  • edit C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v9.2\include\crt\host_config.h Line 131:

from

#if _MSC_VER < 1600 || _MSC_VER > 1913

to

#if _MSC_VER < 1600

OR

  • install vs2017 15.6 from https://docs.microsoft.com/en-us/visualstudio/productinfo/installing-an-earlier-release-of-vs2017

fo40225 avatar Jul 02 '18 01:07 fo40225

Is this wheel(Tensorflow 1.8 with CUDA 9.2)support python2.7? And where should I find a CUDA9.2 wheel which support python2.7?

tgbamg avatar Jul 08 '18 02:07 tgbamg

@tgbamg Uploaded.

https://github.com/fo40225/tensorflow-windows-wheel/tree/master/1.8.0/py27/GPU/cuda92cudnn71sse2

fo40225 avatar Jul 08 '18 06:07 fo40225

no avx I want to second @elranu and give massive props to @fo40225 for compiling the whl file with Cuda 9.2, CuDNN 7.14 and Python 3.6 support in Windows. Your whl file worked like a charm (see my installation environment and hardware details below).

Would you be able to compile a version with AVX support? If so, you would quickly earn yourself the title of "God Status" ;)

Thanks so much for your work - we all appreciate it!

My specs that worked with the WHL:

  • Windows 10 64bit (all latest Windows updates installed)
  • Python 3.6.3 (64bit) [NOTE: I'm just using simple virtual environments with virtualenv v16.0.0 & virtualenvwrapper-win v1.2.5 - NOT Anaconda]
  • Visual Studio Community 2017
  • CUDA Toolkit 9.2.148 (for Windows 10)
  • CuDNN 7.1.4 (for Cuda 9.2)
  • fo40225's compiled WHL: tensorflow_gpu-1.8.0-cp36-cp36m-win_amd64.whl
  • NVIDIA GeForce GTX 1050 Ti with 4gb Ram, NVIDIA Driver version 398.36

jlar0che avatar Jul 31 '18 22:07 jlar0che

@fo40225 Will you build tensorflow-gpu with CUDA 10 support?

RobMcH avatar Sep 21 '18 11:09 RobMcH

I will use CUDA 10 to build tensorflow-gpu 1.11 for python 3.6 x64.

fo40225 avatar Sep 21 '18 11:09 fo40225

I will use CUDA 10 to build tensorflow-gpu 1.11 for python 3.6 x64.

Did you succeed w/above?

bluecd avatar Sep 26 '18 16:09 bluecd

I will use CUDA 10 to build tensorflow-gpu 1.11 for python 3.6 x64.

Did you succeed w/above?

@bluecd Yes.

https://github.com/fo40225/tensorflow-windows-wheel/tree/master/1.11.0/py36/GPU

fo40225 avatar Sep 28 '18 17:09 fo40225