[fix] Add dtype cast for modules other than Transformer

[ir2718]

The PR adds the torch_dtype cast for modules other than Transformer when loading the SentenceTransformer. (#2887)

[ahmedkooli]

Bug reference link :)

[ahmedkooli]

@ir2718 I'm seeing the PR not passing the tests. I think you should change the code to:

if model_kwargs is not None:
    if "torch_dtype" in model_kwargs:
        module = module.to(model_kwargs["torch_dtype"])

[ir2718]

@ahmedkooli

Thanks, for some reason I thought an empty dict was the default.

[tomaarsen]

Heya @ir2718!
Thanks for creating a PR here - I've adopted it with a bit of a different fix, as I noticed some edge cases didn't work:

• {"torch_dtype": "auto"}
• {"torch_dtype": "float16"}
• {"torch_dtype": "bfloat16"}

I've also added a test here. Ideally this should fix #2887

• Tom Aarsen