Gustaf Ahdritz
Gustaf Ahdritz
Since this is a .pdb file, this may be affected by a commit I just pushed (1c3d31b). Re-pull and try again.
I should add that we inherited this OpenMM version number from the original AlphaFold---it's likely that DeepMind ran into the same issue.
Try uninstalling PyTorch from your conda environment and then manually reinstall it using the instructions on the website here: https://pytorch.org/get-started/locally/. LMK if that helps.
Are you sure that your CUDA version isn't actually 10.1?
Not even in your virtual environment anywhere? The error message does confirm that it knows that your torch CUDA is 11.3.
Could you send the output of `pip freeze`? Next, could you try downgrading to torch 1.10.1 and re-running `python3 setup.py install`? We recently upgraded to torch 1.12.0, so that might...
Interesting---this seems to be the first time this is happening on non-Pascal GPUs. I still can't reproduce this @bing-song, so I'll need some extra help here, if you don't mind....
What happens if you put torch.cuda.synchronize() right before that matmul, below the custom kernel call? So strange that the kernel executes multiple times without crashing...
I don't think this has to do with any particular input sequence, since I can't reproduce this on my machine. One last thing, if you don't mind: could you try...