Is pytorch support RTX2080 and cuda10.0 and cudnn7.5?

1

I want to install fastai V1 with pytorch V1

However, I have tried a lot of times, And I cannot get the correct version pytorch.

Here is the technical specs of my PC

GPU: RTx 2080
cuda: 10.0
cudnn: 7.5

here is one piece installation log.
~$ conda install pytorch=1.0.0 cudatoolkit=10.0 -c pytorch
Collecting package metadata: done
Solving environment: /
The environment is inconsistent, please check the package plan carefully
The following packages are causing the inconsistency:

  • defaults/linux-64::anaconda-navigator==1.6.2=py36_0
  • defaults/linux-64::pyqt==5.6.0=py36_2
  • defaults/linux-64::imageio==2.5.0=py36_0
  • defaults/linux-64::seaborn==0.7.1=py36_0
  • defaults/linux-64::pillow==5.1.0=py36h3deb7b8_0
  • defaults/linux-64::qt==5.6.2=h974d657_12
  • defaults/linux-64::qtconsole==4.3.0=py36_0
  • defaults/linux-64::jupyter==1.0.0=py36_3
  • defaults/linux-64::matplotlib==2.2.2=py36h0e671d2_0
  • defaults/linux-64::scikit-image==0.14.2=py36he6710b0_0
  • defaults/linux-64::spyder==3.1.4=py36_0
  • defaults/linux-64::navigator-updater==0.1.0=py36_0
    done

Package Plan

environment location: /home/cine2080/anaconda3

added / updated specs:
- cudatoolkit=10.0
- pytorch=1.0.0

The following packages will be downloaded:

package                    |            build
---------------------------|-----------------
pytorch-1.0.0              |py3.6_cuda9.0.176_cudnn7.4.1_1       498.6 MB  pytorch
------------------------------------------------------------
                                       Total:       498.6 MB

The following NEW packages will be INSTALLED:

pytorch pytorch/linux-64::pytorch-1.0.0-py3.6_cuda9.0.176_cudnn7.4.1_1

2

We’ve had this issue here.
Could you try to run

conda install pytorch=1.0.0 torchvision cuda100 -c pytorch

and see if the right CUDA version will be installed?

2 Likes
3

Screenshot%20from%202019-04-02%2018-03-59
Screenshot%20from%202019-04-02%2018-03-59716×362 41.1 KB

Thanks for your reply!
And here is the screen shot. I think the versions posted are correct.

And still, the cudnn version is not correct

Package Plan

environment location: /home/cine2080/anaconda3

added / updated specs:
- cuda100
- pytorch=1.0.0
- torchvision

The following packages will be downloaded:

package                    |            build
---------------------------|-----------------
cuda100-1.0                |                0           2 KB  pytorch
jpeg-9b                    |       h024ee3a_2         248 KB
pytorch-1.0.0              |py3.6_cuda10.0.130_cudnn7.4.1_1       657.4 MB  pytorch
------------------------------------------------------------
                                       Total:       657.7 MB

The following NEW packages will be INSTALLED:

cuda100 pytorch/linux-64::cuda100-1.0-0
jpeg pkgs/main/linux-64::jpeg-9b-h024ee3a_2
libtiff pkgs/main/linux-64::libtiff-4.0.10-h2733197_2
pytorch pytorch/linux-64::pytorch-1.0.0-py3.6_cuda10.0.130_cudnn7.4.1_1
torchvision pytorch/noarch::torchvision-0.2.2-py_3

I am trying to install it from source.

4

Your system CUDA and cuDNN installations won’t be used if you install the binaries from conda.
The conda bins ship their own packages.
If you want to install from source, have a look here.

5

Did you mean that conda installation will have its own new cudnn and cuda libraries?

And how about the pip method? could you show me the pip install command like the following coding?

conda install pytorch=1.0.0 torchvision cuda100 -c pytorch
6

The pip binaries will also ship their own CUDA and cuDNN libs.
The command should be:

pip install https://download.pytorch.org/whl/cu100/torch-1.0.0-cp37-cp37m-linux_x86_64.whl
7

pip also downloads the binaries for cuda and cudnn. You only need to install cuda and cudnn in your system if you want to compile from source.

Anyway I recommend to use conda. Previously I used Pip until I realized that on some versions numpy was not downloaded. Pytorch documentation recommends to use conda.

1 Like
8

Thanks for your kind help. I use the old command

conda install pytorch=1.0.0 torchvision cuda100 -c pytorch

and with pip install fastai install the pytorch and fastai successfully.

Even though I am not fully understanding the reason why it is cudnn 7.4, but the test result is right.

I test the environment with a mnist example. and here I just show the versions of them


In [2]: import fastai

In [3]: print(torch.__version__)
1.0.0

In [4]: print(fastai.__version__)
1.0.51

And I know the format of my question is very bad but you’d still love to help me our.
Thanks a lot!

Have a good day!

9

Thanks my friend!

I think conda is good. In the future, I will try to use more conda

Actually, I have another question, which is that we can use conda and pip commands install the deep learning toolkit without cuda and cudnn ? Is it ture? or I misunderstand your descriptions?

Have a good day~

10

Well I am not pretty sure. In tensorflow you can just by installing tensorflow or tensorflow-gpu. In Pytorch I have no idea but I think it always download cuda and cudnn binaries.

11

I think the pytorch-cpu binaries won’t ship the CUDA libs and are thus smaller.

1 Like
12

Yes, that is true. From the docs:

conda install pytorch-cpu torchvision-cpu -c pytorch
1 Like