Error on mac gpu with ot.gromov.fgw_barycenters

[youssef62]

Hello,
Is it really necessary to have .double in this line ?
(backend.py line 1944)

  def norm(self, a, axis=None, keepdims=False):
       return torch.linalg.norm(a.double(), dim=axis, keepdims=keepdims)

It has the effect of coverting to float64 which is not supported by mps (gpu for mac silicon) , which raises this error
TypeError: Cannot convert a MPS Tensor to float64 dtype as the MPS framework doesn't support float64. Please use float32 instead.

Is this conversion to float64 intended ?

[rflamary]

I donnt think it is necessary and i don't remember why we have this. This is a recent change

@kachayev you did this change in the backend, do you knwo why? We should try to keep the precision and not change it there.

[rflamary]

We rmoved this conversion in this PR that is now merged.
#572

[youssef62]

Thank you for the quick fix.