Is Taichi able to utilize GPU to speed up the preprocessing within the torch's __getitem__() in the fork mode?

[songyuc]

Hi, guys!
I want to adopt taichi to call my GPU to speed up the preprocessing within pytorch. Since torch cannot use CUDA-tensor in the multi-process loading due to fork start method, I am wondering, whether Taichi can call the GPU in the torch's __getitem__() in this fork mode?

Your answer and guide will be appreciated!