Install Cuda without root

cuda tensorflow gpu theano

30,364

Solution 1

You can install CUDA and compile programs, but you won't be able to run them for a lack of device access.

Solution 2

Update The installation UI for 10.1 changed. The following works:

Deselect driver installation (pressing ENTERon it)
Change options -> root install path to a non-sudo directory.
Press A on the line marked with a + to access advanced options. Deselect create symbolic link, and change the toolkit install path.
Now installation should work without root permissions

Thank you very much for the hints in the question! I just want to complete it with an approach that worked for me, also inspired in this gist and that hopefully helps in situations where a valid driver is installed, and installing a more recent CUDA on Linux without root permissions is still needed.

TL;DR: Here are the steps to install CUDA9+CUDNN7 on Debian, and installing a pre-compiled version of TensorFlow1.4 on Python2.7 to test that everything works. Everything without root privileges and via terminal. Should also work for other CUDA, CUDNN, TensorFlow and Python versions on other Linux systems too.

INSTALLATION

Go to NVIDIA's official release web for CUDA (as for Nov. 2017, CUDA9 is out): https://developer.nvidia.com/cuda-downloads.
Under your Linux distro, select the runfile (local)option. Note that the sudo indication present in the installation instructions is deceiving, since it is possible to run this installer without root permissions. On a server, one easy way is to copy the <LINK> of the Download button and, in any location of your home directory, run wget <LINK>. It will download the <INSTALLER> file.
Run chmod +x <INSTALLER> to make it executable, and execute it ./<INSTALLER>.
accept the EULA, say no to driver installation, and enter a <CUDA> location under your home directory to install the toolkit and a <CUDASAMPLES> for the samples.
Not asked here but recommended: Download a compatible CUDNN file from the official web (you need to sign in). In my case, I downloaded the cudnn-9.0-linux-x64-v7.tgz, compatible with CUDA9 into the <CUDNN> folder. Uncompress it: tar -xzvf ....
Optional: compile the samples. cd <CUDASAMPLES> && make. There are some very nice examples there and a very good starting point to write some CUDA scripts of yourself.
(If you did 5.): Copy the required files from <CUDNN> into <CUDA>, and grant reading permission to user (not sure if needed):

cp -P <CUDNN>/cuda/include/cudnn.h <CUDA>/include/
cp -P <CUDNN>/cuda/lib64/libcudnn* <CUDA>/lib64
chmod a+r <CUDA>/include/cudnn.h <CUDA>/lib64/libcudnn*

Add the library to your environment. This is typically done adding this following two lines to your ~/.bashrc file (in this example, the <CUDA> directory was ~/cuda9/:

export PATH=<CUDA>/bin:$PATH
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:<CUDA>/lib64/

FOR QUICK TESTING OR TENSORFLOW USERS

The quickest way to get a TensorFlow compatible with CUDA9 and CUDNN7 (and a very quick way to test this) is to download a precompiled wheel file and install it with pip install <WHEEL>. Most of the versions you need, can be found in mind's repo (thanks a lot guys). A minimal test that confirms that CUDNN is also working involves the use of tf.nn.conv2d:

import tensorflow as tf
x = tf.nn.conv2d(tf.ones([1,1,10,1]), tf.ones([1,5,1,1]), strides=[1, 1, 1, 1], padding='SAME')
with tf.Session() as sess:
    sess.run(x) # this should output a tensor of shape (1,1,10,1) with [3,4,5,5,5,5,5,5,4,3]

In my case, the wheel I installed required Intel's MKL library, as explained here. Again, from terminal and without root users, this are the steps I followed to install the library and make TensorFlow find it (reference):

git clone https://github.com/01org/mkl-dnn.git
cd mkl-dnn/scripts && ./prepare_mkl.sh && cd ..
mkdir -p build && cd build
cmake -D CMAKE_INSTALL_PREFIX:PATH=<TARGET_DIR_IN_HOME> ..
make # this takes a while
1. make doc # do this optionally if you have doxygen
make test # also takes a while
make install # installs into <TARGET_DIR_IN_HOME>
add the following to your ~/.bashrc: export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:<TARGET_DIR_IN_HOME>/lib

Hope this helps!
Andres

Solution 3

You can install using conda with the following command.

conda install -c anaconda cudatoolkit

But you need to have prior accesss to the device(GPU).

EDIT : If you are finding error in anaconda repository then change the repository to conda-forge which is frequently updated.

conda install -c conda-forge cudatoolkit

30,364

user200340

Updated on July 09, 2022

Comments

user200340 almost 2 years
I know that I can install Cuda with the following:
```
wget http://developer.download.nvidia.com/compute/cuda/7_0/Prod/local_installers/cuda_7.0.28_linux.run
chmod +x cuda_7.0.28_linux.run
./cuda_7.0.28_linux.run -extract=`pwd`/nvidia_installers
cd nvidia_installers
sudo ./NVIDIA-Linux-x86_64-346.46.run 
sudo modprobe nvidia
sudo ./cuda-linux64-rel-7.0.28-19326674.run 
```
Just wondering if I can install Cuda without root?

Thanks,
- Robert Crovella over 7 years
  
  The GPU driver (e.g. sudo ./NVIDIA-Linux-x86_64-346.46.run) is necessary to be able to run CUDA programs on a CUDA GPU, and it requires root privilege to install, the other toolkit components (CUDA toolkit, CUDA samples) can be installed without root privilege, if you direct the installer to place them in your local workspace rather than install to the default locations. If you already have a GPU driver installed on your system that supports the desired CUDA toolkit version then it is possible.
- user200340 over 7 years
  
  @RobertCrovella, looks like I have to ask the admins to install the CUDA installer first, then I can install toolkit myself. Thanks.
- Trect over 5 years
  
  Why dont you use CUDA 10.0
- user200340 over 5 years
  
  @DheerajMPai This question was asked 2 years, 2 months ago. There was no CUDA 10.0 at that time.
lyomi over 6 years

This worked because your machine already had an appropriate driver installed, and all you needed was the CUDA runtime library. If the driver is not installed, there is no way to run the program.
fr_andres over 6 years

that's right, as I say at the beginning I just wanted to complete the informations given by the OP and the accepted answer. I will clarify it
fr_andres over 6 years

In any case installing CUDA without root is a differentiated task from running CUDA, and it is perfectly legitimate to want to install a newer CUDA version on a machine with an existing driver without root permissions...
fr_andres over 6 years

actually, I wanted to edit it but it is already there, in the first sentence. I put it bold so nobody else gets confused
Gochit almost 6 years

@fr_andres: I have a computer with CUDA5. Can I use your tip to install CUDA9 for it?
fr_andres almost 6 years

@Gochit if your OS is debian-based or you can use equivalent commands to do the chmod, wget etc business it should work. Just make sure that your GPU, GPU drivers, CUDA (and optionally CUDNN) versions that you want to install are all compatible among each other. Try to check the GPU and driver version with nvidia-smi, and find the latest version compatible here
fr_andres almost 6 years

Did you try Dheeraj's approach?
Gochit almost 6 years

@fr_andres: I tried your approach, but it reported cudaErrorInsufficientDriver. My driver version is 331.67. I think I can't install tensorflow on this server because the driver is too old. Thank you.
Trect over 5 years

That actually doesn't matter
Ehsan about 5 years

CUDA/10.1 needs root permissions, and I found no way around it. Any suggestions?
fr_andres about 5 years

Just tested the cuda_10.1.105_418.39_linux.run installation and went fine. They changed the UI with some options, please make sure to change the install paths to directories that your user has control of.
fr_andres about 5 years

I updated the answer at the beginning for specific steps, let me know if it works for you!
Trect almost 5 years

Thanks for the elaborate answer.. cp -P <CUDNN>/cuda/include/cudnn.h <CUDA>/include/ cp -P cudnn9/cuda/lib64/libcudnn* cuda9/lib64 chmod a+r cuda9/include/cudnn.h cuda9/lib64/libcudnn* helped a lot
fr_andres over 4 years

@Tessaracter I'm glad to hear that!
szdrnja over 4 years

This is actually a quite straightforward answer but thanks for being thorough! It worked fine for me because I had the drivers but not the libraries I needed for tf.
usamec almost 4 years

That's just libraries. No includes, no compiler, ...
demongolem almost 3 years

On step 4, even after entering my own directory, I got Permission denied. Unable to write to /usr/local/cuda-11.4/ as if it completely ignored my suggestion. I unchecked drivers and left everything else checked. Any suggestions?
fr_andres almost 3 years

It is possible that any errors in the path get silently ignored. Try different paths and combinations, with and without trailing /, no whitespaces... maybe v11 changed for good but hopefully not