SentenceTransformer.encode() moves self (Module) to device

[piteren]

When you call SentenceTransformer.encode() it moves self (Module) to device which by default is None (Line #153 of SentenceTransformer.py). I don't know if it is desired behavior, but I was a bit confused. Imagine such scenario:

model = SentenceTransformer() # I know that I can give device here, while init, but I do not want to
model.to('cuda:1')
model.encode('some text') # it will be encoded on cuda:0 if you have 2 GPUs

[tomaarsen]

I ran into similar issues with moving models to other devices. My solution was to also update model._target_device when moving the model with to.

See the following example:
https://github.com/tomaarsen/setfit/blob/14602ea2773f77b82243624ed1bca5e0772519e7/src/setfit/modeling.py#L442-L445

[tomaarsen]

This confusing behaviour has been updated in #2351, and will be included in the upcoming release.