returnn
returnn copied to clipboard
rwth-i6
Minimum PyTorch version
I think we never officially stated the minimum PyTorch version we require for RETURNN (or did we? where?).
I remember that 1.13 or so was mentioned (or assumed) in some cases but I'm not sure.
Is there a good reason that we don't simply say that we require PyTorch >=2? What is the reason?
E.g., currently I'm thinking about using torch.set_default_device (and torch.device as a context manager) (e.g. in RF PT backend and PT engine) but I think this is only available since PyTorch 2.
What is the reason?
I did not get the hybrid setup to work with 2.0 for various reasons (e.g. Export, Quantization) . Due to the limited time we have for Rescale I wanted to go with 1.13 instead of investigating this further. If someone has time for that to fix this it would be good, but this is beyond my capacity right now.
