Dead Kernel When Importing MXNet (HW1, Q3)

#1

I’m currently in the d2l-en directory on my AWS instance, and able to view the Jupyter interface on my local browser. When I created a new notebook for HW1 (I have been doing all my work on my local machine; will convert over after this homework) and tried to import mxnet, my kernel constantly dies. I can perform simple tasks like printing and addition. I’m running a Python 3 kernel. Any ideas on how to get my import statements to work, and finish #3?

Also, if I try to do it in my Ubuntu machine (running python, then import mxnet as mx), I get an “Illegal instruction (core dumped)” problem.

#2

Could you please give a bit more details? Which GPU? Which version of CUDA? Did you install the NVIDIA drivers, CUDA, CUDNN? There are several libraries that NVIDIA requires and they all come with different install instructions.

#3

NVIDIA-SMI 387.26, CUDA 9.1. I chose the Deep Learning Base AMI Ubuntu 15.0, so I don’t believe I installed any drivers? I followed all the instructions from http://en.d2l.ai/chapter_appendix/aws.html.

#4

Which instance did you launch? Also - have a look at the walkthrough that I posted. It should work just fine. Also, please use CUDA 9.2 instead. I don’t even think that CUDA is installed by default.

ubuntu@ip-172-31-63-98:~$ ls /usr/local
bin  cuda  cuda-10.0  cuda-8.0  cuda-9.0  cuda-9.2  etc  games  include  init  lib  man  python  sbin  share  src
#5

I launched the Deep Learning Base AMI Ubuntu 15.0 instance with a g2.2xlarge GPU.

I followed your walkthrough, and now I’m getting “OSError: libcudart.so.9.1: cannot open shared object file: No such file or directory” when trying to import mxnet as mx. How do I switch Conda versions from 9.1 to 9.2? I have both installed.