I put a tensor data into the GPU through .cuda(), and then use the torch.svd (), but the speed didn’t improve. I found that the CPU was still working at 100%。can’t torch.svd () work in GPU?
We are actively working on adding cuSolver
methods to the torch.linalg
namespace and the PR for svd
is here.
Currently MAGMA would be used, which can fallback to the CPU for smaller sizes of the inputs.