Hi every body.
Does someone know an intuitive workeraround for this issue https://github.com/pytorch/pytorch/issues/5213 ?
Thanks !
I realized the problem is not from torch.split, but rather from the CUDA code for RNN. I’m saying this because I did the operation on CPU using numpy and converted to a tensor of Xx2 I give during the forward pass and it still doesn’t work.
I downgraded to cuda 9.0, cudnn 7.0 and last release of pytorch (0.3.1) and it works.