so 'with torch.no_grad():' couldn't be used in the test?

[WanquanF]

You can use it if you don't need to compute the point normal. If you need the point normal, the grad for the sampled 'u,v' is necessary.