Just wondering when PyTorch will be supporting CUDA 9.x as it seem to be stable now and quite a lot of improvement over the 8.x version.
You can already build pytorch and torchvision to work with CUDA 9
I just got it to work a couple of minutes ago - you can find the install script I am using here, maybe it will be of help.
Thanks for the script !
Link not working, can you repost the script. Also which version of Ubuntu were you on?
I am using 16.04 server since I believe this is the most recent long term of support one? Might be wrong on that one. This is based on the Ubuntu AMI though I am in the process of setting up my own machine - to be completed winter 2020 at this pace
This is the script though I would advise against using it. Pytorch doesn’t work well here - it is slower on a V100 than on the K80? K40? that p2.xlarge is using… I discuss my guesses why it might be in various places on the forum IIRC but it is not worth figuring this out at this point - will wait for when pytorch binaries are available that work with CUDA9. Tensorflow and keras seem to work okay btw but I can imagine only very specific senarios where it would make sense for us to use CUDA9 at this point.
I would suggest the CUDA8 script which I also seem to maintain somewhat and am planning on using it for my own rig once I have all the components in place.