@ttc0419 Exactly. Most of dependencies of transformers
are optional (see repo). By design HuggingFace is aimed at major deep learning frameworks TF, PT, JAX but the issue is that a user usually adhere to a greater extent to a single framework. So, enumeration of only required dependencies gives a user a freedom to manage installed packages on user's system in more fine-grained way.
The errors I listed are NOT optional.
They are actually optional. Just check transformers
repo. It is a nightmare actually. The list of actual dependencies depends on a framework and model. In general it is impossible to manage this kind of projects properly.
Pinned Comments
daskol commented on 2025-04-07 08:09 (UTC) (edited on 2025-04-07 08:09 (UTC) by daskol)
NOTE The next v4.51.0 requires
xet-core
(transitive throughpython-huggingface-hub>=1:0.30.0
). It is now under testing in extra-testing.