I have checked that my tensor is on GPU. The code is too large to be shared here. If you need it, I will try to reproduce it in a small file.
I first want to confirm is there anything else I need to pay attention except adding
ti.cfg.arch = ti.cuda
and putting the tensor and networks to GPU?
BTW, I am on Ubuntu.
Thanks for reporting this issue. I can reproduce it on my end. This will be fix by afternoon.
v0.3.9 supports GPU PyTorch tensors better. Please have a try. (v0.3.9 releasing in 30 minutes.)
If I have used
ti.cfg.arch = ti.cuda, will the tensor automatically be on GPU without explicitly setting device?
No. By default PyTorch places tensors on cpu, and Taichi has to follow this convention…