torch.cuda.is_available() returns False

[tiagoft]

Hello,

I am having trouble using cuda with Pytorch. I am migrating from Theano (maybe this is a problem?). I tried installing pytorch from pip, then uninstalled and tried with conda, and tried compiling from source using python setup install.

It seems that CUDA works everywhere else in the system. I am supposing that this could be a permission issue or an issue with a broken path somewhere? I see that many people are having the same problem; have you been able to solve it somehow?

Thanks!

This is some info about my system and my installation:

• OS: Ubuntu 16.04
• PyTorch version: 0.4.0a0+1ab248d (as stated in torch.__version__)
• How you installed PyTorch (conda, pip, source): from source, although I have tried pip and conda and had the same problem.
• Python version: 2.7
• CUDA/cuDNN version: 8.0 / 7.2.1
• GPU models and configuration: Geforce GTX1070
• GCC version (if compiling from source): 5.4.1

~$ nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2016 NVIDIA Corporation
Built on Sun_Sep__4_22:14:01_CDT_2016
Cuda compilation tools, release 8.0, V8.0.44

~$ nvidia-smi
Tue Mar 27 10:59:19 2018
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 390.42 Driver Version: 390.42 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
|===============================+======================+======================|
| 0 GeForce GTX 1070 Off | 00000000:01:00.0 On | N/A |
| 0% 49C P8 6W / 151W | 266MiB / 8118MiB | 0% Default |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes: GPU Memory |
| GPU PID Type Process name Usage |
|=============================================================================|
| 0 1452 G /usr/lib/xorg/Xorg 144MiB |
| 0 2751 G compiz 118MiB |
+-----------------------------------------------------------------------------+

~$ python -c 'import torch; print torch.cuda.is_available()'
False

~$ python -c 'import torch; print torch.rand(2,3).cuda()'
THCudaCheck FAIL file=/home/username/pytorch/aten/src/THC/THCGeneral.cpp line=70 error=30 : unknown error
Traceback (most recent call last):
File "", line 1, in
RuntimeError: cuda runtime error (30) : unknown error at /home/username/pytorch/aten/src/THC/THCGeneral.cpp:70

I was under the impression that pytorch was not finding CUDA in runtime, so I tried:
~$ CUDA_HOME="/usr/local/cuda" python -c 'import torch; print torch.cuda.is_available()'
False

[ssnl]

What's the relevant log when you compiled from source?

[tiagoft]

I just recompiled again. During the installation, one of the messages was:

-- Found CUDA: /usr/local/cuda (found suitable version "8.0", minimum required is "7.0")

However, I still get:

~$ python -c 'import torch; print torch.cuda.is_available()'
False

@ssnl What log file are you referring to?

[tiagoft]

Ok, problem seems to be solved! How I fixed it (after many, many tries):

1. Installed CUDA 9.1 using apt-get, following the instructions in this link: https://cryptoandcoffee.com/mining-gems/cuda-9-0-install-ubuntu-16-04-apt-get/
2. Installed PyTorch using pip
3. Rebooted

Now I get:
~$ python -c 'import torch; print torch.cuda.is_available()'
True

[StalVars]

"Installed PyTorch using pip", How did you do that? I thought -
PyTorch does not currently provide packages for PyPI (see status at #566).

[imflash217]

@StalVars PyTorch supports both pip and conda package managers.
pip3 install torch torchvision
See https://pytorch.org/

[recolic]

But the source building is still broken, isn't it?

Now I'm suffering from the same problem. Stuck here after fixing many many compilation errors by hand.

[peturhelgi]

I did the same as @tiagoft and it worked. However, I had some problems after logging in as I just got to a blank screen. After reading this thread I resolved the problem.

[bellonemauro]

Hello,

comment here instead of opening a new issue, my device is well configured and works perfectly fine in windows 10 and ubuntu 16.04.
However, I (sometimes) get the error that cuda.is_availble() = false:

RuntimeError: Attempting to deserialize object on a CUDA device but torch.cuda.is_available() is False. If you are running on a CPU-only machine, please use torch.load with map_location='cpu' to map your storages to the CPU.

reboot always solve, but I would like to know what generates this error (I get it ones a day! ) and I have to stop training, save state and reboot.

Thanks
Mauro

[soumith]

hey @bellonemauro , we check with the NVIDIA driver if CUDA / driver is available and ready. If it reports False, we report False. I suspect some weird driver instability.

[bellonemauro]

thank you! this was also my suspect ... no much I can do as I have the latest driver so maybe with some future updates ...

[sammilei]

Hello,

comment here instead of opening a new issue, my device is well configured and works perfectly fine in windows 10 and ubuntu 16.04.
However, I (sometimes) get the error that cuda.is_availble() = false:

RuntimeError: Attempting to deserialize object on a CUDA device but torch.cuda.is_available() is False. If you are running on a CPU-only machine, please use torch.load with map_location='cpu' to map your storages to the CPU.

reboot always solve, but I would like to know what generates this error (I get it ones a day! ) and I have to stop training, save state and reboot.

Thanks
Mauro

reboot doesn't work for me. I have other programs run with CUDA fine but just not pytorch.. (ubuntu 18 cuda 10)

[pyotr777]

Have the same problem.
Installed Pytorch with
pip3 install torch torchvision

Ubuntu 16.04

print(torch.__version__)
1.4.0

$ nvidia-smi
Wed Mar  4 12:07:52 2020       
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 410.78       Driver Version: 410.78       CUDA Version: 10.0     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  Quadro P2000        Off  | 00000000:01:00.0 Off |                  N/A |
| 50%   30C    P0    18W /  75W |      0MiB /  5058MiB |      3%      Default |
+-------------------------------+----------------------+----------------------+
                                                                               
+-----------------------------------------------------------------------------+
| Processes:                                                       GPU Memory |
|  GPU       PID   Type   Process name                             Usage      |
|=============================================================================|
|  No running processes found                                                 |
+-----------------------------------------------------------------------------+

print(torch.cuda.is_available())
False
print(torch.backends.cudnn.enabled)
True

Reboot didn't help.

[nik31096]

See this. It has helped me without reboot every time.

[DarrenCook]

That worked for me. In fact all I needed was:

sudo rmmod nvidia_uvm
sudo modprobe nvidia_uvm

(Trying rmmod nvidia complained about "in use by: nvidia_modeset" - this is on a notebook, and I didn't want to crash the GUI. But it seems it was just nvidia_uvm that needed a kick.)