tensorflow-gpu on Cuda 9.1 Windows 10 Prof

[Ravi Kumar]

Hi,

I pip installed tensorflow-gpu on my laptop with MS windows 10 prof OS successfully.

This is a 8th gen Intel CPU with 16gb Ram with Nvidia Geforce MX150 graphics card.

This graphics card supports only Cuda 9.1 it does not support cuda 9.0.

But when I do " import tensorflow as tf " at the python prompt I am getting error 

" ImportError: Could not find 'cudart64_90.dll' "

and also it is asking me to downgrade to Cuda 9.0

where as in the cuda directory(C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v9.1\bin) I do have  'cudaart64_91.dll'  file .

Any suggestions to overcome this problem ??

Is it hard coded in tensorflow library to look for a certain version of the dll ? 

If not I should be able to provide to change this configuration somewhere in .py file or some txt file without a re-install of tensorflow. 

Regards,

Ravi

[Brett Cuthbertson]

The Windows 10 tensorflow prebuilds 1.5+ are all built against Cuda9.0. That's why it's telling you to downgrade because it can't find cudart64_90.dll in your path. And no, it doesn't work if you just rename cudart64_91.dll to cudart64_90.dll as there are other dependencies.

You've got two choices to get it running

(1) try downloading and installing Cuda9.0 and see if it works anyway. You can have both 9.0 and 9.1 installing concurrently as they install into different directories. I am not sure why your graphics card wouldn't support 9.0?

(2) If you're super keen, compile tensorflow from source and trying building against Cuda9.1, it'll probably work. But you'd have to be super keen. I'd give (1) a try first. 

Brett.

[Ravi Kumar]

Dear Brett Cuthbertson,

 

I appreciate your quick reply.

 

My graphics card does not support cuda 9.0 as you can see at this url : https://devtalk.nvidia.com/default/topic/1023574/cuda-toolkit-9-is-not-available-in-geforce-mx150/

 

So compiling from source is out of question. That is the reason we rely on the binaries provided by tensor flow

 

Regards,

Ravi Kumar

[Toby Boyd]

#2 is likely the best option.  The Nvidia forum seems to suggest that card did not get cuda support until cuda 9.1.  Not saying they are bad, but do these mobile cards do well?  2GB on the card seems small but with smaller batches I imagine it works.  Just curious and maybe derailing the thread.  :-P.  

We will likely be forced to move to 9.2 when it is released.  No promises, forced driver upgrades are unpleasant.  

--
You received this message because you are subscribed to the Google Groups "Discuss" group.
To unsubscribe from this group and stop receiving emails from it, send an email to discuss+unsubscribe@tensorflow.org.
To post to this group, send email to dis...@tensorflow.org.
To view this discussion on the web visit https://groups.google.com/a/tensorflow.org/d/msgid/discuss/6a6ec5a2-1716-4289-80f5-1fab011ebbc1%40tensorflow.org.

[Toby Boyd]

Compiling from source is not out of the question, it is how you would get a 9.1 binary.  Check GitHub in the TF area for CUDA 9.1 and include closed issues.  I started a thread and someone might have made 9.1 windows binaries or have hints.  Cool group of people.

[bd.cuth...@gmail.com]

Just in case you change your mind regarding building from source.

Here are the Windows instructions for building against 9.1

http://www.python36.com/install-tensorflow-gpu-windows/

You could settle in for the night - it'll be more fun than watching an episode of Black Mirror :)

Brett.

To unsubscribe from this group and stop receiving emails from it, send an email to discuss+u...@tensorflow.org.