I've been trying to install the nightly version of pytorch and torchvision because that seems to fix a few things, but only cpu mode works.
Steps
1.a conda create --name night37 python=3.7 create virtualenv (python 3.7)
1.b conda activate night37
1.c load cuda module load cuda10.1
conda install pillow==6.2.1 tqdm ipythonconda install pytorch torchvision cudatoolkit=10.1 --channel pytorch-nightlyipython>>> import torch>>> torch.cuda.is_available()out[2]: FalseUsing Nvidia RTX2080Ti; cuda10.1; RHEL 7, CentOS 7.4.1708
@fmassa this occurs with python 3.7, whereas in python 3.8 torch.cuda.is_available() returns True, but torchvision import breaks resulting in 1827
Binaries were downloaded the 9th of March
Thanks for the bug report!
@seemethere do you know what might be the issue?
This seems fine to me? I cannot reproduce on my local machine
test37 โฏ python -c 'import torch; print(torch.__version__);print(torch.cuda.is_available())'
1.5.0.dev20200310
True
Could you print the output of:
python -c 'import torch; print(torch.__version__);print(torch.cuda.is_available())'
Just made a new virtual env, ssh'd in the compute node and loaded cuda10.1 just to make sure everything was in place.
minimal steps pertaining python are:
conda create --name testenv python=3.7conda install pytorch torchvision cudatoolkit=10.1 -c pytorch-nightlypython -c 'import torch; print(torch.__version__);print(torch.cuda.is_available())'out: 1.3.1, False1.5.0.dev20200309, True so there's a chance the conda channel 'torch-nightly' isn't up to date.In the end, building from source managed to fix both #1827 and pytorch's #4107, which was my main reason to install keira
This looks to be related to the conda uploads failing and our pruner pruning all nightlies available for linux for torchvision.
I have submitted #1960 which should resolve the nightly uploads and will solve this issue.
@mknw Something unrelated: Looks like you didn't activate the new environment and the packages are install in the base environment. Maybe it is just a typo.
@peterjc123 yep, typo
@seemethere thank you, glad I could help!
Thanks @mknw @peterjc123 and @seemethere for the help!