Open scopatz opened 6 years ago
jjhelmus
commented
6 years ago You should not expect cudatoolkit to contain all the files which are included in a full CUDA install. Rather cudatoolkit included the libraries which are redistributable and are used in downstream packages. cuda-install-samples-9.0.sh is not included in the list of redistributable files.
scopatz
commented
6 years ago Thanks for pointing this out @jjhelmus, but I would think that this would fall under section 2.1 of the license which says:
You may distribute the sample source code as delivered by NVIDIA and/or your Derivatives,
provided that all NVIDIA copyright notices and trademarks are maintained and used
properly and the sample source code includes the following notice: “This software contains
source code provided by NVIDIA Corporation.”Also it looks like according to Appendix B, CUDA-GDB seems to be redistributable, because it is GPL'd
jjhelmus
commented
6 years ago Lets back up a bit. The cudatoolkit conda package is intended to provide a common set of shared libraries which are needed by packages such as Tensorflow, PyTorch, XGBoost and CuPy which can make use of CUDA. By providing these libraries in the cudatoolkit package, the host system only needs to provide a compatible NVIDIA CUDA driver to run these various GPU accelerated software packages.
NVIDIA makes available other CUDA related libraries, executables and files. Some of these are available as separate conda packages. For example the NVIDIA CUDA Deep Neural Network (CuDNN) library is available via the cudnn package and the cupti package provides the CUDA Profiling Tools Interface library. Conda packages which make use of these libraries can declare them as dependencies.
There are other components which would be useful to include in conda packages but there are issues with redistribution. An example of this would be, nvcc which would allow run time compilation of CUDA kernels. In theses cases the recommendation is to install these tools directly on the host system.
Finally, there are components of the CUDA offering which could be included in conda packages if there is sufficient interest. My guess is that CUDA-GDB and cuda-install-samples-9.0.sh would be on this list.
It would be good if these components were provided as separate packages rather than including them in the cudatoolkit package itself in order to keep the size of cudatoolkit down and to keep the contents consistent. If you have suggestions for the names, contents or even recipes of these other packages we would welcome them. As a referenc,e the recipe for cudatoolkit can be found in the numba/conda-recipe-cudatoolkit repository.
mikechen66
commented
5 years ago How to delete cuda and cudnn from Conda?
Since Conda-installed CUDA has the error as follows, I want to remove conda-installed cuda and cudnn, and install independent CUDA and cudnn from Nvidia.
Error : Failed to get convolution algorithm. This is probably because cuDNN failed to initialize, so try looking to see if a warning log message was printed above.
However, while I use the commands as follows but can not remove them, I can not remove them.l conda remove --name cuda --all conda remove --name cudnn --all
I see that two documents including cudatoolkit-10.0.130-0 and cudnn-7.3.1-cuda10.0.0_0 in the path as follows.
/home/anaconda3/pkgs/cudatoolkit-10.0.130-0 /home/anaconda3/pkgs/cudnn-7.3.1-cuda10.0.0_0
How can I delete cuda and cudnn embedded in Anaconda.
Thanks in advance,
Mike
Actual Behavior
The example script
cuda-install-samples-9.0.shis not included in the cudatoolkit packageExpected Behavior
This script, and others, should be installed.
Steps to Reproduce
Anaconda or Miniconda version:
Operating System:
conda infoconda list --show-channel-urls