Elle McFarlane

I'm also running into this issue and can reproduce it in an isolated way, e.g.: ``` some_tensor = torch.tensor([1,2,3,4]) print(f"type {type(some_tensor)}, dtype {some_tensor.dtype}") pos_seq = torch.arange(6 - 1, -1, -1.0,...

I fixed it in my fork here and it seems to work now https://github.com/ellemcfarlane/transformers/commit/29739e690bb3448e68914173fea56c1974fba1cb

@amyeroberts > it isn't actively maintained and we won't be making any future changes to the modeling code That's what I figured, so no worries :)

Hi! I'm still seeing this as an issue for latest version: image-reward 1.5