We're still using it in @spacy_io . But we don't want to try to make efficient transformer architectures in it. We've always called out to other libraries for that. This lets us wrap PyTorch implementations directly, which makes things a bit easier for us.

→ View original post on X — @honnibal,