Runtime error 999 when trying to use cuda with pytorch

user_10396945 0
Viewed 14306

I installed Cuda 10.1 and the latest Nvidia Driver for my Geforce 2080 ti. I try to run a basic script to test if pytorch is working and I get the following error:

RuntimeError: cuda runtime error (999) : unknown error at ..\aten\src\THC\THCGeneral.cpp:50

Below is the code im trying to run:

import torch
torch.cuda.current_device()
torch.cuda.is_available()
torch.cuda.get_device_name(0)
pytorch python
4 Answers

Restarting my computer fixed this for me.

But for a less invasive fix, you can also try this solution (from a tensorflow issue thread): 

sudo rmmod nvidia_uvm
sudo rmmod nvidia
sudo modprobe nvidia
sudo modprobe nvidia_uvm
user_72111370

In case of PyTorch it seems to be sufficient to restart only nvidia-uvm (Unified Virtual Memory) with

sudo modprobe --remove nvidia-uvm  # same as `rmmod`
sudo modprobe nvidia-uvm

If that doesn't work, go ahead and restart the whole module with additional modprobe (--remove) nvidia as mentioned @matwilso's answer.

user_26876010

In my case, I solved the 999 error with: nvidia-modprobe -u

user_16455450

Reinstalling NVIDIA driver solves this problem in my case. (You don't need to restart the computer)

user_79789030
Related