[fix] Added model2vec import compatible with current and newer version

[Pringled]

Hi!

Pull Request overview

• Make sentence-transformers compatible with current and newer version of Model2Vec

Details

Starting from model2vec version 0.3, inference and distillation will be split up since inference will only need Numpy. This makes it much faster to initialize and install, since there is no dependency on Torch anymore for inference. This PR makes sentence-transformers compatible with both the old and new version, including an updated ImportError message for from_disillation.

Relevant Model2Vec commit: MinishLab/model2vec@010a7b7

[tomaarsen]

Hello!

This looks solid I think! Also, torch-less inference on model2vec itself is quite nice, well done.

Have you tested from_model2vec with the latest model2vec? I think this could fail:

sentence-transformers/sentence_transformers/models/StaticEmbedding.py

Lines 202 to 203 in 1802076

	static_model = StaticModel.from_pretrained(model_id_or_path)
	embedding_weights = static_model.embedding.weight

• Tom Aarsen

[tomaarsen]

I added some simple tests to make sure that things don't break in the future.

[Pringled]

@tomaarsen good catch, those need to be converted! Will update the code

[tomaarsen]

@Pringled I already have a fix locally, I'll take care of it :)

[Pringled]

Nice! That should work with both the old and new versions 💪