I was running the following code on GPU:
import torch X = torch.rand(4, 4) X = X.cuda() torch.cumsum(X, dim=0)
and I got this error:
THCudaCheck FAIL file=/home/shuoyangd/pytorch/torch/lib/THC/generic/THCTensorMathScan.cu line=52 error=8 : invalid device function
What confuses me is that other functionalities (matrix multiplication etc.) seem to work fine (on the same GPU). So just to confirm, is this just that
torch.cumsum function does not have GPU support yet?
BTW, I’m using Tesla K80.