When I use these functions on Cuda tensors I have 0% GPU usage. Are math functions computed on the CPU regardless of the fact that tensors are on the GPU?
I believe some parts run on the GPU and some parts run on the CPU. Are you really getting 0% GPU usage?
Yes, I’m monitoring it with
nvidia-smi --loop=1 so it is not very accurate. I get about 7% in the first second, then it goes to zero.
Hi, I have the same problem as you, have you solved your problem?