When running 3dgs on a machine with multiple graphics cards, it reports memory at dist2 = torch.clamp_min(distCUDA2(torch.from_numpy(np.asarray(pcd.points)).float().cuda()), 0.0000001) # (P,)

[p-z-p]

trafficstars

You can see that the memory is only used for a very small portion, and there is no problem running the same data on a single gpu machine, why is that?