In the past I wrote everything myself in c++ / cuda. So I have seen all this mention of handy libraries for ML. So I thought I would see what it was all about.
So I follow the steps on PyTorch website. I install Anaconda.
First problem, no instructions on what command to install pytorch via Annaconda. just some vague reference to a command you need to run.
So I google it and find a random website with a script to do it!! https://deeplizard.com/learn/video/UWlFM0R_x6I
wtf guys, anyway the script is an old version of Cuda.
conda install pytorch torchvision cudatoolkit=10.0 -c pytorch
so I try a current version
conda install pytorch cudatoolkit=11.1 -c pytorch
Next problem:
Anaconda complains that “The following packages are not available from current channels”
cudatoolkit=11.1
so thanks for wasting my time Pytorch. I have no way to use you with the current cuda libraries. So back to doing it myself.
Then you simply choose your method, for example I’d like to use pip and install the cuda11 version, I click on pip, then cuda 11.0 and bingo I get my commands to install pytorch :
Its up to the Team Pytorch @ptrblck may have a better answer as to why.
It maybe that the Cuda11 was released back in May, and the codebase was tested on it. Cuda11.1 was released around september 23 I guess, around 1 month before releasing the stable 1.7.
So it makes sense to stick to the version you have been working on for the past couple of months and not make yourself deal with some probable regression/bugs caused by the new update. (take into account the OSes, drivers, etc as well)
If there is a need for bleeding edge cuda, anyone can easily build from source and call it a day. For the majority of users though the stability is the primary goal.
Nevertheless, we will be expecting a 1.7.1 in a near future (possibily in December?) and that might include cuda 11.1 I’m not sure though.
Again as I stated earlier, you can always build Pytorch against the latest cudatoolkit and enjoy it. the build process is pretty easy to follow and usually is very straightforward.
The binaries are not shipping with CUDA11.1, as we couldn’t prune libs with this version and were running into errors. If you need to use the latest library versions (CUDA, cudnn, NCCL etc.), you could always build PyTorch from source or use the NGC docker containers.
Your A100s work with CUDA11.0 and since sm86 is SASS binary compatible with sm80, also the 30XX series works.