CUDA: Check failed: e == cudaSuccess (999 vs. 0) : unknown error

Hello Forum,

I have a problem and I still have it for approximately one or two weeks and I’m getting crazy a little bit… I hope someone of you could help to solve my problem…
Okay, first my environment, I created a docker container with Ubuntu 20 (docker image name, ubuntu:latest). I installed cuda and cuda-drivers, I installed Jupyter for development and debugging, I compiled Apache MxNet configured for Jetson Xavier NX, after all these steps I can register the Apache MxNet module in Python. The registration was successful and I can execute the version command. Also, I can use mxnet with CPU support BUT, here is my problem, I canNOT use it with GPU support. And exactly this is what I like to do…

Nvcc output:
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2021 NVIDIA Corporation
Built on Sun_Feb_14_21:13:57_PST_2021
Cuda compilation tools, release 11.2, V11.2.152
Build cuda_11.2.r11.2/compiler.29618528_0

Python version:
Python 3.8.5

Apache MxNet version output:
‘1.8.0’

Always when I insert this source code inside of a Jupyter Notebook I get the error message written in the title and in the following. I tried a lot of things, for instance, I though the Apache MxNet compilation failed because I used wrong parameters, so I tested a lot of different parameters but it still does not work… I installed all dependencies again and again but it still does not work…

Source Code:
import mxnet as mx
mx.test_utils.list_gpus()

Error Message:
MXNetError: Traceback (most recent call last):
File “include/mxnet/base.h”, line 459
CUDA: Check failed: e == cudaSuccess (999 vs. 0) : unknown error

Config.mx (last compilation):

I hope someone of you have an idea for me to solve this problem, I would be very glad to solve this!!

Thank you very much for every help!

Julian M.

Hi,

Not sure if we understand your use case correctly.

There is a host (x86) platform that try to compile a MXNet for Xavier with docker.
If yes, please noted that you will need a cross compiler to generate the ARM package.

If you can launch the package on an x86 environment, it indicates the package doesn’t compile for ARM.
And you will need to add the local dGPU architecture to enable it on an host environment.

More, we have a pre-compiled MXNet package in our eLinux wiki for your reference:

https://elinux.org/Jetson_Zoo#MXNet

Thanks.

Hey AastaLLL,

thank you very much for your answer.

I tried it and at the end of the script, does not matter which either autobuild_mxnet.sh or autoinstall_mxnet.sh, an error occurs:

When I try to import Apache MxNet in Python3 I get the same error.

Do you have an idea for a solution?

Thank you.

Here is another screenshot with the result of the script autoinstall_mxnet.sh

now I tried an older version, the installation seems to be okay but the result is still the same…

Hello,

I solved my problem! Thank you very much!

I used this link for solving the problem “Illegal instruction (core dumped)”:

And I used this link for installation:

Thanks for the update!