glearn.device.locate_cuda#

glearn.device.locate_cuda()#

Returns the directory paths and version of CUDA Toolkit installation.

Note

Either of the environment variables CUDA_HOME, CUDA_ROOT, or CUDA_PATH should be set with the home directory of the CUDA Toolkit installation before calling this function.

Returns:
cudadict

A dictionary with the following keys:

  • home: str, home directory of CUDA.

  • nvcc: str, the path to nvcc compiler file.

  • lib: str, the path to the library directory of CUDA.

  • include: str, the path to the directory of header files.

  • version: dict, the version of CUDA Toolkit with the following keys:

    • major: int, the major number of version

    • minor: int, the minor number of version

    • parth: int, the patch number of version

If no CUDA Toolkit is found, it returns an empty dictionary {}.

Raises:
EnvironmentError

Raised if in the home directory of the CUDA Toolkit, the expected sub-directories cannot be found.

See also

glearn.info
glearn.device.get_nvidia_driver_version

Notes

Setting Environment Variables:

In order to find CUDA Toolkit information properly, either of the environment variables CUDA_HOME, CUDA_ROOT, or CUDA_PATH should be set to the directory where CUDA Toolkit is installed. Usually on UNIX operating systems, this path is /usr/local/cuda. In this case, set CUDA_HOME (or any of the other variables mentioned in the above) as follows:

export CUDA_HOME=/usr/local/cuda

To permanently set this variable, place the above line in profile file, such as in ~/.bashrc, or ~/.profile, and source this file, for instance by

source ~/.bashrc

Note

It is possible that the CUDA Toolkit is installed on the machine, but cuda_version key shows not found. This is because the user did not set the environment variables mentioned in the above.

Expected Directory Structure:

This function looks for the executable nvcc (or nvcc.exe if windows) in the directory specified by CUDA_HOME, CUDA_ROOT or CUDA_PATH environment variables. If nvcc executable is found, it continues searching for the directory structure as described below.

The expected sub-directories under the CUDA home directory should have the following structure:

  • /bin with the executable /bin/nvcc (or /bin/nvcc.exe in Windows)

  • /include with the file /include/cuda.h

  • /lib or lib64 in UNIX, and lib/x86 or lib/x64 in Windows.

Examples

Suppose the CUDA Toolkit is installed in /usr/local/cuda. First, in a Shell environment, export the variable

export CUDA_HOME=/usr/local/cuda

Then in Python script, obtain info about the CUDA paths and version using:

>>> # Import function
>>> from glearn.device import locate_cuda
>>> cuda = locate_cuda()

>>> # Neatly print the dictionary using pprint
>>> from pprint import pprint
>>> pprint(cuda)
{
    'home': '/usr/local/cuda/11.7',
    'include': '/usr/local/cuda/11.7/include',
    'lib': '/usr/local/cuda/11.7/lib64',
    'nvcc': '/usr/local/cuda/11.7/bin/nvcc',
    'version':
    {
        'major': 11,
        'minor': 7,
        'patch': 0
    }
}