Hi,
I am trying to install pytorch via anaconda in Ubuntu 20.04 with CUDA 11.
However, I didn’t find the installation option for CUDA 11 on the “Get started” webpage.
Does that mean I have to go back to CUDA 10.2?
Thx.
3 Likes
I am also facing the same dilemma. Let me know, if, you figured out a workaround solution.
1 Like
me too.
If not yet available, when planning to support CUDA 11?
The binaries are not built yet and you would have to install PyTorch from source at the moment.
2 Likes
@ptrblck
I successfully build PyTorch from source with CUDA 11. However, I still got:
➜ pytorch python
Python 3.8.2 (default, Jul 16 2020, 14:00:26)
[GCC 9.3.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import torch
>>> torch.cuda
<module 'torch.cuda' from '/home/jiapei/.local/lib/python3.8/site-packages/torch/cuda/__init__.py'>
>>> torch.cuda.is_available()
False
>>> print(torch.version.cuda)
11.0
>>> torch.backends.cudnn.enabled
True
>>>
Clearly, the output of torch.cuda.is_available() returned False, which is incorrect.
Any further suggestions?
Your driver might be too old, as CUDA11 needs >= 450.36.06.
1 Like
⋊> ~ nvidia-smi 09:56:01
Wed Aug 19 09:56:04 2020
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 450.57 Driver Version: 450.57 CUDA Version: 11.0 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|===============================+======================+======================|
| 0 GeForce GTX 105... Off | 00000000:01:00.0 Off | N/A |
| N/A 37C P8 N/A / N/A | 1439MiB / 4042MiB | 17% Default |
| | | N/A |
+-------------------------------+----------------------+----------------------+
Could you post the install log, so that we can have a look, if CUDA was successfully detected?
Hi, @ptrblck
In fact, even now, I’m NOT sure how to install PyTorch correctly. For me, I’ve got to build twice, one using cmake, the other python setup.py build etc.
Under ./build/CMakeFiles, there are 2 .log files:
Any suggestions?
Unfortunately, I cannot access the files. The install log would be still helpful.
The build instructions can be found here.
I tried running this command conda install pytorch==1.6.0 cudatoolkit=11.0 -c pytorch with python v3.7.10, Nvidia-driver 450.119.03 and CUDA version 11.0.
But the problem is torch.cuda.is_available() is returning False and torch.cuda.version is returning None.
It showed me the following logs:
added / updated specs:
- cudatoolkit=11.0
- pytorch==1.6.0
The following packages will be downloaded:
package | build
---------------------------|-----------------
cudatoolkit-11.0.3 | h15472ef_8 951.9 MB conda-forge
------------------------------------------------------------
Total: 951.9 MB
The following NEW packages will be INSTALLED:
cffi conda-forge/linux-64::cffi-1.14.5-py37hc58025e_0
cudatoolkit conda-forge/linux-64::cudatoolkit-11.0.3-h15472ef_8
future conda-forge/linux-64::future-0.18.2-py37h89c1867_3
libblas conda-forge/linux-64::libblas-3.9.0-8_mkl
libcblas conda-forge/linux-64::libcblas-3.9.0-8_mkl
liblapack conda-forge/linux-64::liblapack-3.9.0-8_mkl
llvm-openmp conda-forge/linux-64::llvm-openmp-11.1.0-h4bd325d_1
mkl conda-forge/linux-64::mkl-2020.4-h726a3e6_304
ninja conda-forge/linux-64::ninja-1.10.2-h4bd325d_0
numpy conda-forge/linux-64::numpy-1.21.0-py37h038b26d_0
pycparser conda-forge/noarch::pycparser-2.20-pyh9f0ad1d_2
pytorch conda-forge/linux-64::pytorch-1.6.0-cpu_py37hf1c21f6_1
typing_extensions conda-forge/noarch::typing_extensions-3.10.0.0-pyha770c72_0
The following packages will be DOWNGRADED:
_openmp_mutex 4.5-1_gnu --> 4.5-1_llvm
Proceed ([y]/n)? y
Any suggestions?
1 Like
You are installing the CPU version as shown in the logs, since 1.6.0 didn’t ship with cudatoolkit=11.0.
3 Likes
Yeah, I had to install pytorch 1.7.0 which worked fine.
Command I used:
conda install pytorch==1.7.0 cudatoolkit=11.0 -c pytorch
I meet the same problem ,just using mmdetection, the terminal shows : The detected CUDA version (11.0) mismatches the version that was used to compile
PyTorch (10.2). Please make sure to use the same CUDA versions.
To build custom CUDA extensions your local CUDA toolkit will be used and should match the used CUDA runtime shipped in the pip wheels / conda binaries. Install the CUDA11 wheels and it should work.
Thank you! I just want to learn the relation of the cuda，pytorch，cudann，cudatoolkit,where can I see that?
You can print the different versions using print(torch.__config__.show()) and would see the used CUDA runtime, cuDNN etc.
Although the original question is on Ubuntu, I had this problem on windows 11.
nvidia-smi
Tue May 24 21:51:38 2022
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 516.01 Driver Version: 516.01 CUDA Version: 11.7 |
|-------------------------------+----------------------+----------------------+
| GPU Name TCC/WDDM | Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|===============================+======================+======================|
| 0 NVIDIA GeForce ... WDDM | 00000000:01:00.0 Off | N/A |
| N/A 33C P5 17W / N/A | 0MiB / 6144MiB | 0% Default |
| | | N/A |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=============================================================================|
| No running processes found |
nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2022 NVIDIA Corporation
Built on Tue_May__3_19:00:59_Pacific_Daylight_Time_2022
Cuda compilation tools, release 11.7, V11.7.64
Build cuda_11.7.r11.7/compiler.31294372_0
Here is how I solved the problem:
- I updated Anaconda : I switched to the latest version, and rebooted my computer after the installation.
- I chose the last version of pytorch here
conda create -n global_env
conda activate global_env
conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch
import torch
print(torch.cuda.is_available(), torch.version.cuda, torch.backends.cudnn.enabled)
True 11.3 True
torch.__version__
‘1.11.0’
Hi @ptrblck ,
I am trying to install pytorch via cmd in windows 10 with CUDA 11.4.2
How to install pytorch which is compatible with this CUDA version?
Thanks
We did not build PyTorch binaries shipping with CUDA 11.4, if I’m not mistaken. You could install the latest PyTorch release with CUDA 11.8 or 12.1 instead.