Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Planning Ticket CUDA 9.2 + cuDNN 7.1 #18906

Closed
tfboyd opened this issue Apr 26, 2018 · 57 comments

Comments

Projects
None yet
@tfboyd
Copy link
Member

commented Apr 26, 2018

Updated 02-JUNE-2018

I have done some testing with CUDA 9.2/cuDNN 7.1.4/NCCL 2.x with tf_cnn_benchmarks.

Below you will see that if you upgrade to cuDNN 7.1.4 and a newer device driver you get some pretty nice gains and no need to compile from source as well as using CUDA 9.0. Compiling from source and using NCCL 2.x (Hierarchal Copy was almost the same) I was able to get 6,800+ 8xV100s ResNet50v1 FP16 with synthetic data and about 6,500+ real data sustained over a full run for under 5 hours.

I suspect many of you saw NVIDIA announce 1K and 1.3K for 1xV100 ResNet50. That was with unreleased libraries and we are working to ensure we can hit those numbers when the libraries are available or with our own tricks.

I apologize for any short hand I use below in describing the runs. I am happy to answer questions or make clarifications. This testing was slightly informal but recent full scale testing yielded similar results.

CUDA 9.0
[Recommended driver] 6,197.70 CUDA 9.0 + 384.13 (v1.8.0-1386-g2dc7575) hierarchical copy
6,620 CUDA 9.0 + 390.59 (v1.8.0-1386-g2dc7575) NCCL
6,613.11 CUDA 9.0 + 396.26 (v1.8.0-2215-gf528eba) NCCL
6,564.9 CUDA 9.0 + 396.26 (v1.8.0-2215-gf528eba) hierarchical copy
6,541.25 CUDA 9.0 + 390.59 (1.9.0.dev20180523) hierarchical copy
6,276.02 CUDA 9.0 + 384.13 (1.9.0.dev20180523) hierarchical copy
6,197.70 CUDA 9.0 + 384.13 (v1.8.0-1386-g2dc7575) hierarchical copy

CUDA 9.1
[Recommended driver] 6,227.64 CUDA 9.1 + 390.59 (v1.8.0-2215-gf528eba) hierarchical copy
6,210.28 CUDA 9.1 + 396.26 (v1.8.0-2215-gf528eba) hierarchical copy
6,118.21 CUDA 9.1 + 396.26 (v1.8.0-2215-gf528eba) NCCL

CUDA 9.2
[Recommended driver] 6,696.39 CUDA 9.2 + 396.26 (v1.8.0-2215-gf528eba) NCCL
6,606.57 CUDA 9.2 + 396.26 (v1.8.0-2215-gf528eba) hierarchical copy
6,738.32 CUDA 9.2 + 396.26 (v1.8.0-2215-gf528eba) NCCL SGD

Full test runs with CUDA 9.2
top_1 ranges between 75.7% and 76% and does not seem to be based on the hyper parameters. I was focused on testing NCCL vs. hierarchical copy. I have a minor concern about my validation command, but this is still good info.
Hierarchical copy: 6441.64 Accuracy @ 1 = 0.7584 Accuracy @ 5 = 0.9267 [49920 examples]
NCCL (repacking:2): 6490.62 Accuracy @ 1 = 0.7572 Accuracy @ 5 = 0.9265 [49920 examples]
NCCL (repacking:8): 6490.62 Accuracy @ 1 = 0.7582 Accuracy @ 5 = 0.9268 [49920 examples]
My first hierarchical copy run was 76% exactly.

Note: I would like to move to CUDA 9.2 as the default but the driver is not widely available for easy apt-get install. I am working long-term to get the build team to support a secondary CUDA build. I am also very open to feedback as I do not see any easy path forward to moving to newer CUDA versions faster. You can always (usually) install a new cuDNN. We are compiling with 7.0 but you see gains by installing 7.1.4 as you see above if you upgrade your driver.

@tensorflowbutler

This comment was marked as off-topic.

Copy link
Member

commented Apr 27, 2018

Thank you for your post. We noticed you have not filled out the following field in the issue template. Could you update them if they are relevant in your case, or leave them as N/A? Thanks.
Have I written custom code
OS Platform and Distribution
TensorFlow installed from
TensorFlow version
Bazel version
CUDA/cuDNN version
GPU model and memory
Exact command to reproduce

@SephirothFFKH

This comment has been minimized.

Copy link

commented Apr 29, 2018

Great! I was able to build TF with the configuration below but the biggest roadblock is to get a version of TensorRT that works with anything above 9.0 hopefully TensorRT 4.1 comes right in time for this r1.9

OS Platform and Distribution: Lubuntu 17.04
TensorFlow installed from Source
TensorFlow version r1.8
GCC version 7
Bazel version 0.12
CUDA/cuDNN version 9.1/7.1
NCCL 2.1.15
TensorRT None
GPU model and memory Dual GTX1080Ti + GTX980Ti

@ViktorM

This comment has been minimized.

Copy link

commented May 11, 2018

When 1.9 release can be expected, at least approximately?

@tfboyd

This comment has been minimized.

Copy link
Member Author

commented May 11, 2018

1.9 had a code cut off of 31st May. It normally takes 1-2 weeks to verify the RCs and let them "bake". CUDA 9.2 will not be in 1.9 as the default due to CUDA 9.2 not being public yet and in my opinion the device driver is too new and still not available for Debian although you can get it for Ubuntu as "beta". I am hopeful NVIDIA is making some changes that will reduce this pain.

@SephirothFFKH I am just getting more involved in including TensorRT. I did the testing for the NVIDIA "launch" but that was very one off. Thank you for the info on versions working with 9.2.

@ViktorM

This comment has been minimized.

Copy link

commented May 11, 2018

@tfboyd Thanks for clarification! So 1.9 will be shipped with CUDA 9.0 or at least switch to 9.1 could be expected?

@alanpurple

This comment has been minimized.

Copy link

commented May 17, 2018

CUDA 9.2 GA is just released and support official driver of course
FYI

@K3n4

This comment has been minimized.

Copy link

commented May 17, 2018

I can't find cudnn for cuda 9.2 windows

@achalshah20

This comment has been minimized.

Copy link
Contributor

commented May 17, 2018

I am trying to build tensorflow with cuda 9.2. Although cudnn doesn't support cuda 9.2 officially.

@alanpurple

This comment has been minimized.

Copy link

commented May 18, 2018

@K3n4
@achalshah20

cudnn also just been released

cudnn 7.1.4 for cuda 9.2

@achalshah20

This comment has been minimized.

Copy link
Contributor

commented May 18, 2018

@alanpurple Thanks!

@K3n4

This comment has been minimized.

Copy link

commented May 18, 2018

@alanpurple Thanks,
Is it possible to build tf with this configuration ?

@alanpurple

This comment has been minimized.

Copy link

commented May 19, 2018

@K3n4
I'll skip that one until 1.9rc

as a developer it's not a good timing for trying those, since 1.8 has many compiling issue( intel MPI support for example)

and 1.9 will come with prebuilt binary for cuda 9.2 and cudnn 7.1(so compiling will be promising), and mpi-support issue fixed

1.9rc are planned to come within 2 weeks as far as I know

so I recommend you also wait for 1.9rc

@mirekphd

This comment has been minimized.

Copy link

commented May 19, 2018

You've missed CUDA 9.1 and now decide to simply skip it? People will never be able to switch from CUDA 8.0 if the support for CUDAs 9.x among ML algos is so randomly scattered among these 3 versions (and counting)...

@mirekphd

This comment has been minimized.

Copy link

commented May 19, 2018

Yes, Tensorflow 1.7 can be built against CUDA 9.1 and CuDNN 7.1.2 if you have 1-2 hours per each try and don't want automated builds (e.g. for Docker). Here's how to do it, dear Tensorflow team.

@mirekphd

This comment has been minimized.

Copy link

commented May 19, 2018

But wait, there's more: TinyMind prepared a comprehensive set of pre-compiled Tensorflow wheels, including those for CUDA 9.1. Make sure to star their hard work.

@Harbing

This comment has been minimized.

Copy link

commented May 19, 2018

I was not able to install CUDA 9.1 on Ubuntu 16.04-64bit a week ago.
I just managed to have CUDA 9.2 running on Ubuntu 16.04-64bit, and cuDNN 7.1.4 that is compatible with CUDA 9.2.
However, Tensorflow GPU does not seem to work with CUDA 9.2. It keeps complaining "ImportError: libcublas.so.9.0: cannot open shared object file: No such file or directory".
Any ideas?

@alanpurple

This comment has been minimized.

Copy link

commented May 20, 2018

@mirekphd no i'm using 9.1 to build tf-gpu 1.8 and i'm using it
i'm skipping 9.2 for 1.8, waiting 1.9rc

@mirekphd

This comment has been minimized.

Copy link

commented May 20, 2018

@alanpurple, if you have a working wheel for TF 1.8 compiled under CUDA 9.1 and CuDNN 7.1.2, then consider contributing it to TinyMind, they seem to be missing 1.8

@mirekphd

This comment has been minimized.

Copy link

commented May 20, 2018

@Harbing:
I had a similar error message caused by Tensorflow expecting CUDA 9.0, when I tried to import it in a CUDA 9.1 NVIDIA-docker container. What solved the problem for me was the obvious thing: using a different build of TF, compatible with my CUDA 9.1 version. I used the CUDA 9.1 Tensorflow 1.7 Python 3.6 wheel from TinyMind. It loaded correctly and passed our usability test under CUDA Version 9.1.85 and cuDNN 7.1.2 (i.e. in the latest NVIDIA CUDA 9.1 container - nvidia/cuda:9.1-cudnn7-devel).

@fredguth

This comment has been minimized.

Copy link

commented May 24, 2018

@mirekphd, unfortunately there is no CUDA 9.2 build there. To sum up, the problem is that Nvidia is pointing users to install CUDA 9.2, but the available builds are 9.0 compatible only. It seems that the only solution is either downgrade to CUDA 9.0 (could use some help here, I couldn't find a simple way in Nvidia's site) or to compile Tensorflow from source.

@purpledawn777

This comment has been minimized.

Copy link

commented Jun 3, 2018

...well I take that back. after exiting the command line and opening a new command window and re-activating the conda environment I did the installation in, going into python no longer showed the error above.
running python without pre-activating the conda environment did complain.
installation successful.

@VRscience

This comment has been minimized.

Copy link

commented Jun 3, 2018

@purpledawn777 Are you using CUDA 9.2 toolkit? No other CUDA toolkit is installed in your machine?
I am still struggling with it but I do not really see how to solve the issue but downgrade to CUDA 9.0...

@purpledawn777

This comment has been minimized.

Copy link

commented Jun 4, 2018

@VRscience I am using toolkit 9.2
I installed in an anaconda environment and I first created a tensorflow conda environment, as instructed in the tensorflow instructions.
when testing it complained about not finding cuda 9.0.
I then installed the wheel nicely provided by fo40225
when I tested immediately it still had the error, but then when I reopened the command line, activate tensorflow and tested again, it worked.... sorta -- it now complained about missing cuDNN7, but I just had to install that from Nvidia's website into the cuda 9.2 directories and it works fully now.

@VRscience

This comment has been minimized.

Copy link

commented Jun 4, 2018

@purpledawn777 I did exactly the same but i keep on having the same issue. Also, it seems there is no more need of installing the wheel provided by @fo40225 as on (tensorflow)C:> pip install --ignore-installed --upgrade tensorflow-gpu the wheel installed is already tensorflow_gpu-1.8.0-cp36-cp36m-win_amd64.whl .

Any idea?

@tfboyd

This comment has been minimized.

Copy link
Member Author

commented Jun 5, 2018

@mirekphd
Sorry for the slow response. Every PR for TensorFlow is unit tested across every platform we support and then internally run against an even large set of tests. Adding another version of CUDA expands the matrix and we are currently using hundreds of GPUs just for the unit testing much less performance testing. CUDA 9.1 ended up being a slight perf regression depending on how you look at the results and what is compared. I like all the builds people do to try out new things.

NVIDIA's official DGX-1 also was not upgraded and still, as of today, only supports CUDA 9.0 with the official image.

@fo40225 You likely figured this out but you can just use the CUDA 9.0 cuDNN with CUDA 9.1. That is what NVIDIA suggested when I saw that a few days back. I would not bother with CUDA 9.1.

@ViktorM 1.9 will be CUDA 9.0. CUDA 9.1 seems to have a perf regression (check out my updated original post above) or I might have pushed for the upgrade after the driver for CUDA 9.2 was newer than I expected. I am still working on NCCL 2.x as that provides a nice boost, but there are some license issues to work out.

Note: I think it is cool other people post how to compile. The community likely does a better job than I would as I take short cuts and have some weird preferences as to how I like to install things. I also only do linux and old python (@patrikhuber :-)) What I can do in the future is post how I compiled along with the sha-hash or branch where I did so. Far from perfect and if other people are ahead of me that is great. I have a small insight. TensorFlow is part of the Brain team at Google. While some might upgrade to the latest versions of cuDNN and such, it is not something I hear people talking about. I am still hopeful that CUDA 10 will bring a new approach to drivers that makes it easier for us to roll out minor CUDA bumps with less pain.

As a random note. I strongly suggest using Docker. I use to deal with all of these different versions of CUDA in python virtual environments. I feel kind of silly having not moved to docker sooner. So much better.

@yukoba

This comment has been minimized.

Copy link

commented Jun 5, 2018

If someone wants to install CUDA 9.2 to Ubuntu 18.04, use this.

sudo apt-key adv --fetch-keys http://developer.download.nvidia.com/compute/cuda/repos/ubuntu1710/x86_64/7fa2af80.pub
echo "deb https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1710/x86_64 /" | sudo tee /etc/apt/sources.list.d/cuda.list
sudo apt update
sudo apt -o Dpkg::Options::="--force-overwrite" install cuda-9-2 cuda-drivers
sudo reboot
@tfboyd

This comment has been minimized.

Copy link
Member Author

commented Jun 5, 2018

I still think all of you are nuts (meant to be funny) not to just use docker and just keep your local device driver updated. :-) That is how I do the performance testing, if you were curious. Total docker convert here and I fought it to the end.

@mtianyan

This comment has been minimized.

Copy link

commented Jun 6, 2018

I build the cuda9.2 for TensorFlow1.8.0 on python3.6 use the ubuntu18.04

https://github.com/mtianyan/tensorflow-linux-wheel

@alanhsieh2000

This comment has been minimized.

Copy link

commented Jun 8, 2018

I migrated to TensorFlow r1.8 with cuda 9.2.88+patch 1, cuDNN 7.1.4 from r1.3 20 days ago. Maybe some issues I experienced and found the way out would help you.
#19371

@XiaoqiChai

This comment has been minimized.

Copy link

commented Jun 8, 2018

@alanhsieh2000
Hi, I also have cuda 9.2.88 with patch1, cuDNN 7.1.4, and tensorflow-gpu 1.8 in conda virtue environment with python 3.6. But I got following problem. Do you have any ideas?

ImportError: libcublas.so.9.0: cannot open shared object file: No such file or directory

Thanks!

@alanhsieh2000

This comment has been minimized.

Copy link

commented Jun 8, 2018

@XiaoqiChai
Did you build tensorflow r1.8 from source? It looks like your python is trying to load cuda 9.0 rather than cuda 9.2. This information is specified when you configure Bazel before you started to build. You have to give it 9.2 rather than 9, when the config tool asks you for the cuda version.

@hadim

This comment has been minimized.

Copy link

commented Jun 11, 2018

I have recently created a Dockerfile that make it easy for anyone to compile tensorflow in Linux: https://github.com/hadim/docker-tensorflow-builder

@Sri06006

This comment has been minimized.

Copy link

commented Jun 18, 2018

Install Cuda 9.2.88_windows 10, and cuDNN 7.1 for cuda 9.2. dll is copied to "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v9.2\bin".
installed tensorflow gpu python 3.6
But getting error while importing tensorflow
(tf15) C:\Users\91966>python
Python 3.6.5 |Anaconda, Inc.| (default, Mar 29 2018, 13:32:41) [MSC v.1900 64 bit (AMD64)] on win32
Type "help", "copyright", "credits" or "license" for more information.

import tensorflow
Traceback (most recent call last):
File "C:\Users\91966\Anaconda3\envs\tf15\lib\site-packages\tensorflow\python\platform\self_check.py", line 75, in preload_check
ctypes.WinDLL(build_info.cudart_dll_name)
File "C:\Users\91966\Anaconda3\envs\tf15\lib\ctypes_init_.py", line 348, in init
self._handle = _dlopen(self._name, mode)
OSError: [WinError 126] The specified module could not be found

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
File "", line 1, in
File "C:\Users\91966\Anaconda3\envs\tf15\lib\site-packages\tensorflow_init_.py", line 24, in
from tensorflow.python import pywrap_tensorflow # pylint: disable=unused-import
File "C:\Users\91966\Anaconda3\envs\tf15\lib\site-packages\tensorflow\python_init_.py", line 49, in
from tensorflow.python import pywrap_tensorflow
File "C:\Users\91966\Anaconda3\envs\tf15\lib\site-packages\tensorflow\python\pywrap_tensorflow.py", line 30, in
self_check.preload_check()
File "C:\Users\91966\Anaconda3\envs\tf15\lib\site-packages\tensorflow\python\platform\self_check.py", line 82, in preload_check
% (build_info.cudart_dll_name, build_info.cuda_version_number))
ImportError: Could not find 'cudart64_90.dll'. TensorFlow requires that this DLL be installed in a directory that is named in your %PATH% environment variable. Download and install CUDA 9.0 from this URL: https://developer.nvidia.com/cuda-toolkit

I tried to downgrade to 9.0, but failed to so.

Is there a solution for issue#1 and if not how do I downgrade to 9.0?

@Enumaris

This comment has been minimized.

Copy link

commented Jun 25, 2018

@VRscience Have you found a solution to your issue yet? I am getting the exact same issue (and it looks like @Sri06006 has run into this issue as well). It does look like the wheel being installed when running pip install --upgrade tensorflow-gpu is the one that purports to be for cuda 9.2 (the one @fo40225 posted) and yet when you look into the build_info after installation, it definitely says the build version is for cuda 9.0. So I am confused. I've tried uninstalling and reinstalling, clearing my cache and then reinstalling, and nothing seems to work. The only thing I can think to do right now is to install cuda 9.0 alongside cuda 9.2 (they can coexist afaik) and see if that helps, but it seems like I will be missing out on any potential performance boosts with cuda 9.2.

@Sri06006

This comment has been minimized.

Copy link

commented Jun 26, 2018

problem is resolved. Have found solution .
could not point to the right one. Here is what I did
Uninstalled all CUDA files in my system, cleared the registry

Installed CUDA 9.0 and unchecked the visual studio integration during installation
and copied the dll from cuDNN to bin folder in v9.0/
you can try installing cuda 9.0 and set the environment Path to
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v9.0\bin

Restart the computer after setting the environment variables

install the tensorflow-gpu

@Enumaris

This comment has been minimized.

Copy link

commented Jun 26, 2018

@Sri06006 I was also able to get things working by installing CUDA 9.0. I installed it along side CUDA 9.2 since I was told they can coexist. Things are working, but I still think it's odd that the whl file's description would say it has CUDA 9.2 support if it only explicitly looks for CUDA 9.0.

@beew

This comment has been minimized.

Copy link

commented Jun 29, 2018

I have successfully built and run tensorflow1.8 and 1.9rc1 against cuda9.2+patch1 and cudnn 7.1, with python3.5 on Ubuntu 16.04. I installed cuda9.2 stuffs in a separate test folder (use the .run files without sudo)

I source this script when building and whenever I run these versions of tf (tf1.9 compiled with openmpi. but need to change line 76 of tensorflow/tensorflow/contrib/mpi_collectives/kernels/mpi_ops.cc from se to stream_executor or build would fail)

$PYTHONUSERBASE is set to the test folder so pip3 install --user would install the test tf whl (only one of 1.8 or 1.9rc can exist of course)inside the test folder without messing up the system's version. To invoke it would need to prepend $PYTHONPATH accordingly.

This way it would invoke the test version of tf and it would point to the matching version of cuda (9.2 instead of system's 9.1)

export PREFIX=/home/beew/opt/cuda_test/cuda92
export PATH=$PREFIX/cuda/bin:$PREFIX/bin:$PATH
export CUDA_SDK_ROOT_DIR=$PREFIX/samples/common
export TENSORRT_PATH=$PREFIX/TensorRT-4.0.1.6

export LD_LIBRARY_PATH=$PREFIX/cuda/lib64:$PREFIX/cuda/extras/CUPTI/lib64:$LD_LIBRRAY_PATH:$TENSORRT_PATH/lib

export PYTHONUSERBASE=$PREFIX

export PYTHONPATH=$PREFIX/lib/python3.5/site-packages:$PYTHONPATH

export MPI_HOME=/usr/lib/openmpi

export CPATH=$PREFIX/include:$CPATH
export LIBRARY_PATH=$PREFIX/lib:$LIBRARY_PATH
export LD_LIBRARY_PATH=$PREFIX/lib:$LD_LIBRARY_PATH

alias nvblas92="LD_PRELOAD=$PREFIX/cuda/lib64/libnvblas.so"
@tensorflowbutler

This comment has been minimized.

Copy link
Member

commented Jul 14, 2018

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

@alanpurple

This comment has been minimized.

Copy link

commented Jul 15, 2018

this issue can be closed, it's working with cuda 9.2 and cudnn 7.1 already

@sliedes

This comment has been minimized.

Copy link

commented Jul 21, 2018

@tfboyd I'm sorry that this will probably sound like a rant, but since you explicitly solicited for feedback...

You say that you want to cater for a wide audience and therefore not force an upgrade to a new driver, and I'd say that's the right approach. However, in my experience due to the combination of (1) TensorFlow requiring a specific version of CUDA, and (2) distributions generally not supporting having multiple distributor-provided CUDA versions simultaneously installed, TensorFlow is one of the most difficult pieces of software to get to work. Almost every time I have installed TensorFlow on a computer there has been a mismatch of CUDA/cublas/cudnn/whatever versions. This is a major hassle, to the extent that usually my first reaction to code that uses TensorFlow is to search for an implementation for one of the other deep learning frameworks (I do have a CPU-only version of TensorFlow installed on my development workstation currently). I just don't want to go through the hassle of maintaining manually installed parallel versions of CUDA. The fact that you praise Docker as a solution should be good evidence that things are not as easy as they should – though probably it's still good advice.

Right now, a competing deep learning framework doesn't yet support CUDA 9.2, but I find that forgivable when they provide binaries for CUDA versions 8, 9.0 and 9.1, with 9.2 coming soon. I don't know how they manage it, but installing the other framework has never been a hassle, while installing TensorFlow has always been painful.

I do think however (with little actual knowledge) that Linux distributions could provide coinstallable CUDA packages (like cuda-9.0, cuda-9.1, cuda-9.2), just like there can be multiple different-sonamed versions of libraries. I can't understand why it's not done like that; at least for the required libraries it seems the exact same problem.

So, if the goal is to cater for a wide audience, sorry to say but in my opinion it's not working.

@MaisaDaoud

This comment has been minimized.

Copy link

commented Jul 27, 2018

after thousands of trails which included adding the variables to the paths, installing cuda9.0 besides cuda9.2 and installing cudnn7.0 from the archive nothing solved the problem except when I restarted my machine. Not sure if restarting would solve the problem without the other steps

@tfboyd

This comment has been minimized.

Copy link
Member Author

commented Jul 31, 2018

@sliedes The part you may be missing is licensing and the legal ability to distribute something like CUDA, cuDNN (officially it is behind a login), or the other associated libraries.

As a side thing, I updated the install guide to include installing everything from NVIDIA with apt-get. I have done this myself on a clean system and it seems to be smooth. I would still do it from docker as it is easy to get the packages mixed up and get the versions misaligned.

On a positive side CUDA 10 will not require device driver updates and amusingly enough is going to be more backwards compatible than CUDA 9.2. I and I am sure others have been working with NVIDIA to simplify everything.

@tfboyd tfboyd closed this Jul 31, 2018

@fangyihao

This comment has been minimized.

Copy link

commented Aug 2, 2018

Hopefully, this wheel file below helps.
TensorFlow 1.9.0 build with cuda 9.2, cudnn 7.1.4 and python 3.5 on ubuntu 16.04:
https://github.com/fangyihao/tensorflow-wheels/blob/master/tensorflow-1.9.0-cp35-cp35m-linux_x86_64.whl

@LukeSBE

This comment has been minimized.

Copy link

commented Sep 13, 2018

hi is there a Windows build that works with Cuda 9.2?

9.0 is 10 months out of date and tensorflow is unusable on windows. do people only update tensorflow once a year?

@haramoz

This comment has been minimized.

Copy link

commented Sep 26, 2018

Does someone have a way to build tensorflow-gpu for centOS and cuda 9.2. My cluster admin has updated driver and removed cuda 9.0 i can not fin anything for cuda 9.2 and CENTOS Linux. If anyone can make one or have a clue please let me know. Thank you :(

@beew

This comment has been minimized.

Copy link

commented Sep 26, 2018

I have been able to build tensorflow-gpu on Cuda 9.2 and Cuda 10 on Ubuntu 16.04, basically just configure, choose your options and install whatever missing. Shouldn't be a problem on Centos if you have up to date bazel and Nvidia driver

@tfboyd

This comment has been minimized.

Copy link
Member Author

commented Oct 3, 2018

I have started a thread on CUDA 10.

#22706

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
You can’t perform that action at this time.