Runtime issues with VGGT-SLAM 2

[AlexMorgand]

Hey thank you for this great upgrade as I enjoyed VGGT SLAM 1 already :)

I'm trying to run the default office sequence with the vggt slam 2 but getting some issues:

1. We now use half precision for vggt but seems like the forward is getting an issue with this "RuntimeError: Input type (float) and bias type (c10::BFloat16) should be the same" when doing "predictions = model(images)".
2. Commenting this for now, it's running when doing loop closure we have "TypeError: VGGT.forward() got an unexpected keyword argument 'compute_similarity'" at line 399 in the solver.py "predictions_lc = model(lc_frames, compute_similarity=True)"

I also had to comment "#line_width=3.0," and doing "image=img/255.0," for the add_camera_frustum probably as a consequence that we don't use float16?

Thank you in advance!

[Dominic101]

@AlexMorgand can you check that you are using our fork of VGGT which should have been installed by setup.sh? If you go into thirdparty/vggt, run git log and the most recent commits should be from me. Also check that this is the installation of vggt being used and not another local installation