globalATenDispatch c++ memory exception after upgrading to 1.2 stable

Hi all,

I am using libtorch and I’ve just upgraded from Pytorch 1.1 to 1.2 stable. The (formerly working code) compiles ok, but I cannot even torch::cat two GPU tensors because of a c++ memory exception in

static inline Tensor cat(TensorList tensors, int64_t dim)

somewhere around globalATenDispatch.

From what I undestand, this dispatcher is something new, what am I missing here? Should the functions like cat be somehow registered to the dispatcher first?

Running Win 10.



@Alex1 This seems like a bug (similar to Could you post an issue at We will triage properly.

Ok, I’ve created an issue, it’s easily reproducible.