Training is already done in PyTorch. CUDA is for low level optimizations.

→ View original post on X — @cerebras