Consistent use of torch device in ‎HitAndRunPolytopeSampler

[mhrmsn]

Issue description

In botorch.utils.sampling.‎HitAndRunPolytopeSampler.draw (see here), when calling sample_polytope, the tensors A and b are explictly moved to the CPU, but not x0.

Is this intended, and if so why? If not, should this be changed to moving x0 also explicitly to the CPU or by taking the source devices of all tensors?

x0 seems to ultimately come from the equality constraints (x0 -> self.nullC -> Vh -> self.C -> equality_constraints). Depending on where the equality constraints come from (external library that calls botorch) this might be an issue, and given that those tensors must anyway be on the same device for multiplication in sample_polytope, it might make sense to change this.

System Info

Please provide information about your setup, including

• BoTorch Version (0.11.3)
• GPyTorch Version (1.12)
• PyTorch Version (2.3.0)
• macOS Sonoma, 14.6.1

[Balandat]

Is this intended, and if so why?

No, I don't think so. x0 is constructed as a new tensor there, so it will be constructed on the default torch device. If that is not the cpu then this could cause errors. It would make sense to explicitly construct this tensor on the GPU, let me put up a quick PR.

[mhrmsn]

Is this intended, and if so why?

No, I don't think so. x0 is constructed as a new tensor there, so it will be constructed on the default torch device. If that is not the cpu then this could cause errors. It would make sense to explicitly construct this tensor on the GPU, let me put up a quick PR.

Perfect, thanks!