It seems there is a problem with the conda installation of the 1.0.0 version for cuda 10. If I execute:
/usr/local/anaconda3/bin/conda install -y pytorch=1.0.0 torchvision cudatoolkit=10.0 -c pytorch
Then If executing. I get cuda version 9:
(/tmp/cuda10.0_pytorch_1.0.0/) user@usermachine1:/tmp$ python -c "import torch; print(torch.version.cuda)"
Could you try to install it using
cuda100 instead of
conda install pytorch=1.0.0 torchvision cuda100 -c pytorch
Okei, three months ago it was in the way you mention, two months ago in the way I have done, and now it seems it has changed again. Do you now which will be the final syntax?
I’m not sure, if it changed for
cudatoolkit was introduced in
1.0.1 (release notes).
Well, I did an installation of several different versions of pytorch in february on some machines and need to change from cu100 to cudatoolkit=10.0 (even for other cuda version or older pytorch as 0.4.0). Maybe they have change it again.
Yes, maybe. As far as I know Soumith and Paul (not sure how his user name is here) are creating the conda builds, so they should know what’s been changed and provide more information.
Okei. I have checked my git repo and the scripts I used to install several versions of Pytorch by February 2019 looked like (for python 2.7, 3.5 and 3.7):
$conda install -y pytorch=1.0.0 torchvision cudatoolkit=10.0 -c pytorch
$conda install -y pytorch=0.4.0 torchvision cudatoolkit=9.0 -c pytorch
and so on. So things might have changed from then. In fact, I initially used your syntax but by that time that did not work. I mean:
$conda install -y pytorch=1.0.0 torchvision cuda100 -c pytorch
$conda install -y pytorch=0.4.0 torchvision cuda90 -c pytorch
Moreover, it seems strange that by typing cudatoolkit=10.0 with version 1.0.0 what I get installed is cuda9. Moreover, some versions cannot be installed with conda and only with pip:
-python 3.5 cuda9 version 0.4.0
-python2.7 cuda9 version 0.4.0.
As far as I checked
cudatoolkit version should be the final version with no planned changes. It’s what the Anaconda team recommended, and we worked with them to move to this.