Pytorch_geometric: Cuda 11 and RTX 30 Series GPU Support

Created on 24 Oct 2020 · 6Comments · Source: rusty1s/pytorch_geometric

❓ Questions & Help

Is there going to be a new version that supports Cuda 11 above and RTX 30 Series GPUs soon?

Source

johnny12150

Most helpful comment

CUDA 11 should be supported when manually installing from source. We will officially support it once PyTorch 1.7.0 lands.

rusty1s on 25 Oct 2020

👍3

All 6 comments

CUDA 11 should be supported when manually installing from source. We will officially support it once PyTorch 1.7.0 lands.

rusty1s on 25 Oct 2020

👍3

Should be any day now. For some reason I seem to have errors with using .jit when combined with forward/dataparallel, same network with conv2d doesn't fail :/ Hopefully python 1.7.0 fixes this.

aluo-x on 26 Oct 2020

Really look forward to CUDA 11 support.

kunshouout on 30 Oct 2020

I was under the impression that CUDA 11 had some bugs (I think slow performance on tensor cores for 3D conv), that were fixed in CUDA 11.1 (or the new cudnn?).

There's a bunch of similar bugs in the pytorch bug tracker.

aluo-x on 30 Oct 2020

It seems that PyTorch 1.7.0 has landed.

johnny12150 on 7 Nov 2020

I just upgraded my dev environment to CUDA 11.1/cudnn 8/pytorch 1.7.0.
Aside from having to modify the magma build scripts, I was able to install pytorch + pytorch geometric without additional patches.

aluo-x on 7 Nov 2020

Was this page helpful?

0 / 5 - 0 ratings