"conda install cudatoolkit" failed on Sagemaker studio lab - Can't create "lib/libnppif.so"

Question

“conda install cudatoolkit” failed on Sagemaker studio lab with the following error: 
InvalidArchiveError("Error with archive /home/studio-lab-user/.conda/pkgs/cudatoolkit-11.8.0-h37601d7_11hobyfh16/pkg-cudatoolkit-11.8.0-h37601d7_11.tar.zst.  You probably need to delete and re-download or re-create this file.  Message from libarchive was:

Can't create 'lib/libnppif.so'")

We have tried deleting the archive file "/home/studio-lab-user/.conda/pkgs/cudatoolkit*" and reinstalled cudatoolkit but still received the same error.

Can anyone share where "lib/libnppif.so" (a standard CUDA shared lib) is included the default Sagemaker studio lab environment so that we can run the "conda install cudatoolkit" command.  Or let us know where the standard cuda library paths are on Sagemaker studio lab (nothing obvious based on our search) so that we can try "EXPORT LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda/lib64"

Thank you!

Answer

[CUDA Toolkit](https://developer.nvidia.com/cuda-toolkit). Before you do `conda install cudatoolkit`, please make sure that you are using a SageMaker Studio lab runtime with GPU enabled. I tried to duplicate your issue on my side, I can hardly get a GPU enabled runtime. Base on my experience, I would recommend you to set up a SageMaker Studio domain with instance type of ml.p3.2xlarge to perform experiemental task.

"conda install cudatoolkit" failed on Sagemaker studio lab - Can't create "lib/libnppif.so"

Contenuto pertinente