Having trouble finding libcuda.so.1 file despite CUDA being installed.

[saya18]

I've installed CUDA and have a Nvidia driver installed but when I run a Python module that uses CUDA, I'm getting an
error that "libcuda.so.1 file" is not found.

I'm using ubuntu 16.04 and using Python 3.6 /Python 2.7

This is the exact error:


ImportError: libcuda.so.1: cannot open shared object file: No such file or directory



From what I've been told, this file is a link and it needs to be linked to my graphics card driver.

I wasn't sure how to check if i have this link but I ran this command and I got back this as an output (abridged):


find / -type f -name "libcuda.so.1





ind: ‘/etc/cups/ssl’: Permission denied
find: ‘/etc/polkit-1/localauthority’: Permission denied
find: ‘/etc/ssl/private’: Permission denied
find: ‘/tmp/systemd-private-bfc953f066c54c8f8989b0585e58681d-colord.service-QhckWW’: Permission denied
find: ‘/tmp/systemd-private-bfc953f066c54c8f8989b0585e58681d-systemd-timesyncd.service-A46ooI’: Permission denied
find: ‘/tmp/systemd-private-bfc953f066c54c8f8989b0585e58681d-rtkit-daemon.service-pZ6U3J’: Permission denied
find: ‘/lost+found’: Permission denied
find: ‘/var/tmp/systemd-private-7216baf4e9e24f4b99aa9cd9d37e9779-rtkit-daemon.service-vEpGYO’: Permission denied
find: ‘/var/tmp/systemd-private-c9508c53c88848febd8d6b9c7758d44d-colord.service-6sVMbw’: Permission denied
find: ‘/var/tmp/systemd-private-7216baf4e9e24f4b99aa9cd9d37e9779-systemd-timesyncd.service-DifcXc’: Permission denied
find: ‘/var/tmp/systemd-private-7216baf4e9e24f4b99aa9cd9d37e9779-colord.service-j5hYyg’: Permission denied
find: ‘/var/tmp/systemd-private-81dcc732570e47799cb04c3cb0c5a2c6-systemd-timesyncd.service-dSg1Cz’: Permission denied
find: ‘/var/tmp/systemd-private-f72e80f0374645bda6c2d99c5628e374-colord.service-FbxlSK’: Permission denied
find: ‘/var/tmp/systemd-private-5065912711c44bfd880f3aca2d0008e7-colord.service-rq0MKq’: Permission denied
find: ‘/var/tmp/systemd-private-bfc953f066c54c8f8989b0585e58681d-rtkit-daemon.service-W2mqTy’: Permission denied
find: ‘/var/tmp/systemd-private-5065912711c44bfd880f3aca2d0008e7-rtkit-daemon.service-Nmhoc5’: Permission denied
find: ‘/var/tmp/systemd-private-bfc953f066c54c8f8989b0585e58681d-colord.service-yD6AKb’: Permission denied
find: ‘/var/tmp/systemd-private-310aa08f8dac48c087fb3d04eb13211d-rtkit-daemon.service-2aRSdk’: Permission denied
find: ‘/var/tmp/systemd-private-cc0e6bd6ee4c4e5a8e66d39c662b4262-systemd-timesyncd.service-cR7tKn’: Permission denied
find: ‘/var/tmp/systemd-private-81dcc732570e47799cb04c3cb0c5a2c6-colord.service-RpnOff’: Permission denied
find: ‘/var/tmp/systemd-private-93e35b4b8e084692829998454c625032-rtkit-daemon.service-FPP0C0’: Permission denied
find: ‘/var/tmp/systemd-private-f72e80f0374645bda6c2d99c5628e374-rtkit-daemon.service-KSb7II’: Permission denied
find: ‘/var/tmp/systemd-private-93e35b4b8e084692829998454c625032-colord.service-umcrrr’: Permission denied





Running this command:


ls /usr/local/cuda-8.0/doc/man/man7/libcuda.so.7 -la


Got me this output:


-rw-r--r-- 1 root root 26 Jan 26 2017 /usr/local/cuda-8.0/doc/man/man7/libcuda.so.7


But I couldn't find libcuda.so.1 within the same vicinity of folders.

If i run:

And if I run nvidia-smi I get back this:





------------------------------------------------------+
| NVIDIA-SMI 340.104 Driver Version: 340.104 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
|===============================+================= =====+======================|
| 0 GeForce GTX 260 Off | 0000:01:00.0 N/A | N/A |
| 40% 46C P12 N/A / N/A | 226MiB / 895MiB | N/A Default |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Compute processes: GPU Memory |
| GPU PID Process name Usage |
|================================================= ============================|
| 0 Not Supported







I'm stumped as to how to overcome this error.

Is libcuda.so.7 also a symlink?

How do I find where libcuda.so.1 is ?
And how do I link it with my graphic card driver (I don't know how to locate the driver as well)

Thank you.

[monkeybrain20122]

What is the output for

locate libcuda.so.1

[Yellow Pasque]

Do you have packages libcuda1-340 and pyhton3-pycuda installed?

[saya18]

What is the output for

locate libcuda.so.1


Hi. I seem to get back an empty response.

Does that mean libcuda.so.1 is not installed?


EDIT:

If I browse my computer directory, I noticed that there is a CUDA-8.0 at this location:


/usr/local/

I also see a CUDA folder with an arrow on it (which I presume represents a shorcut) here in


/usr/local/

[saya18]

Do you have packages libcuda1-340 and pyhton3-pycuda installed?


I'm not sure what python3-pycuda is, but I try "locate libcuda1-340" I get back a blank response.

[deadflowr]

libcuda would be in the /usr/lib/x86_64-linux-gnu directory (or /usr/lib/i386-linux-gnu if on 32-bit Ubuntu)
How'd you install cuda?

[saya18]

libcuda would be in the /usr/lib/x86_64-linux-gnu directory (or /usr/lib/i386-linux-gnu if on 32-bit Ubuntu)
How'd you install cuda?


It was a few weeks ago so it's either this command:

sudo apt install nvidia-cuda-tookit

Or I ran a run file which I downloaded from the Nvidia website.


Using this command adds the path to the environmental variable?


export PATH=$PATH:/usr/local/cuda-8.0/bin

I just tried it and now nvcc seems to work.

Now when I run
nvcc --version







nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2016 NVIDIA Corporation
Built on Tue_Jan_10_13:22:03_CST_2017
Cuda compilation tools, release 8.0, V8.0.61




However, when I run


locate libcuda.so.1

I get back an empty screen.

[mc4man]

If you installed your nvidia drivers from a .deb package then a corresponding libcuda package would also be installed. If you install the drivers from a .run file then I gather libcuda may come with it (the .7 .so) but apparently you need one named .1

So you could probably
1. Remove the current nvidia drivers & install from a .deb package (- making sure you get at least the min. version needed for your cuda toolkit version
2. create a symlink named libcuda.so.1 linking to the actual .so you have.

More current nvidia drivers in .deb form can be gotten here
https://launchpad.net/~graphics-drivers/+archive/ubuntu/ppa

[pauljw]

I'm no expert here at all, but I have sometimes needed to: "sudo updatedb" before the locate command would show recently installed files. Maybe that's the case here.

[saya18]

If you installed your nvidia drivers from a .deb package then a corresponding libcuda package would also be installed. If you install the drivers from a .run file then I gather libcuda may come with it (the .7 .so) but apparently you need one named .1

So you could probably
1. Remove the current nvidia drivers & install from a .deb package (- making sure you get at least the min. version needed for your cuda toolkit version
2. create a symlink named libcuda.so.1 linking to the actual .so you have.

More current nvidia drivers in .deb form can be gotten here
https://launchpad.net/~graphics-drivers/+archive/ubuntu/ppa


Going through my terminal commands it seems I installed Nvidia drivers like this:

sudo apt-get install nvidia-current
sudo add-apt-repository
sudo add-apt-repository ppa:graphics-drivers
sudo apt-get update

sudo apt-get install nvidia-340

This seems to be same method you are suggesting?

Thank you.


I remember I was having problems with with some versions, but nvida-340 seems to work fine.

[saya18]

I'm no expert here at all, but I have sometimes needed to: "sudo updatedb" before the locate command would show recently installed files. Maybe that's the case here.


Thank you. I tried it, but still can't find the file.

[pauljw]

No problem, thought it may be worth a shot. :)

[The Cog]

It is probably worth taking mc4man's suggestion and creating a symlink named libcuda.so.1 to your existing file. The later version probably just has new features that anything wanting to use version 1 won't call.

[saya18]

It is probably worth taking mc4man's suggestion and creating a symlink named libcuda.so.1 to your existing file. The later version probably just has new features that anything wanting to use version 1 won't call.


Thank you. So is this the proper to create a symlink in my case?



ln -s /usr/local/cuda-8.0/doc/man/man7/libcuda.so.7 libcuda.so.1

[The Cog]

I always have trouble with paths and symlinks. I would do it this way:

cd /usr/local/cuda-8.0/doc/man/man7/
sudo ln -s libcuda.so.7 libcuda.so.1
If it doesn't work, you can just rm the symlink again.

[saya18]

I always have trouble with paths and symlinks. I would do it this way:

cd /usr/local/cuda-8.0/doc/man/man7/
sudo ln -s libcuda.so.7 libcuda.so.1
If it doesn't work, you can just rm the symlink again.


Thank you. I tried what you suggested, but I'm still getting the same error:


ImportError: libcuda.so.1: cannot open shared object file: No such file or directory




It seems like I did create the symlink, as If I run the code, I get that there is a libcuda.so.1 file already.
So is the next problem a path issue?

Not sure what I should be adding to the PATH.

Thank you.

[The Cog]

Hmm. I wonder if it needs to be in a different location. If so, I don't know where it needs to be.

[Yellow Pasque]

/usr/local/cuda-8.0/doc/man/man7/libcuda.so.7

This is a man(ual) page. This is not what you need. Symlinking to a manpage will not help you.


"locate libcuda1-340" I get back a blank response.

No. libcuda1-340 is the package which has the file you need. Make sure it is installed:

sudo apt-get install libcuda1-340
Is there a reason you couldn't use the version in the repository or at least use Nvidia's .deb file?

[saya18]

This is a man(ual) page. This is not what you need. Symlinking to a manpage will not help you.



No. libcuda1-340 is the package which has the file you need. Make sure it is installed:

sudo apt-get install libcuda1-340
Is there a reason you couldn't use the version in the repository or at least use Nvidia's .deb file?


Ah, I see. Going to try installing it. I tried the .debian file, but I forget the problems I was having. Honestly, I tried so many ways to install CUDA TOOLS and Nvidia drivers, that I lost count.
I was following the Nvidia instructions, then multiple blogs and get running into strange bugs (login bug with newest drivers). I was finally able to install drivers and CUDA (which I thought) but following instructions on some blog.
So that's why my install setup has been so wacky.


UPDATE:

I installed libuca1-340,

and now I seem to be getting a different error, not sure if it's because of the earlier symlink I created or maybe a PATH issue?

ImportError: /home/moondra/.local/lib/python3.6/site-packages/tensorflow/python/../libtensorflow_framework.so: undefined symbol: cuDevicePrimaryCtxSetFlags

[mc4man]

see here
https://github.com/tensorflow/tensorflow/issues/6787
The gist of it seems to be to try newer nvidia driver than 340, the ppa has 384, 387, 390, ubuntu repos for 16.04 probably has 384

[Yellow Pasque]

see here
https://github.com/tensorflow/tensorflow/issues/6787
The gist of it seems to be to try newer nvidia driver than 340, the ppa has 384, 387, 390, ubuntu repos for 16.04 probably has 384

The two people who reported success there were using a GTX750 (Maxwell) and a Grid K520 (Kepler), so they had the option to run the latest stable driver.
OP here is using a GTX 260, and is limited to 340.xx driver.

[Yellow Pasque]

TensorFlow requires "GPU card with CUDA Compute Capability 3.0 or higher"
https://www.tensorflow.org/install/install_linux

Your GTX 260 only supports CUDA Compute 1.3:
https://developer.nvidia.com/cuda-legacy-gpus

You must use the CPU version of TensorFlow (or get a better GPU).

[saya18]

TensorFlow requires "GPU card with CUDA Compute Capability 3.0 or higher" https://www.tensorflow.org/install/install_linux Your GTX 260 only supports CUDA Compute 1.3: https://developer.nvidia.com/cuda-legacy-gpus You must use the CPU version of TensorFlow (or get a better GPU). Ah. Thank you so much. Despite, listing my graphics card info, none of the nvdia members pointed this out. I thought because my graphics card was CUDA okay, I could use it with Tensorflow etc. Anyway, after restarting the computer I'm stuck in a login loop, which I will make another thread about.