The infer speed on the RTX 4060 is slow.

[jiang-junlong]

Hi, thank you for great work, I wonder know why the speed on the RTX 4060 laptop is very low, about 0.8s.
Is there any other way to increase the inference speed? I have installed the correct version of xformers.

[camaosos]

the torch_cluster.knn method runs slow in GPU. I had the same problem. My workaround was to move everything to cpu only for that method

[finnBsch]

Hi
Agree with @camaosos. The reason for CPU being faster seems that it can then use KDTrees. I have created a PR (see #26) that allows you to always use KDTrees if you want to, and it will use torch_kdtree, which also requires CPU to build the tree but then runs inference on GPU, which should be faster than full CPU.

[jiang-junlong]

Thanks very much, I will test today.