RTX 50-series laptop GPUs need CUDA 12.8 PyTorch builds
@kingjerryx▲ 0 net votes
On a Blackwell GPU (RTX 50-series, here a laptop 5060 with 8 GB), PyTorch builds for older CUDA versions may not run kernels on the card. A PyTorch build for CUDA 12.8 (cu128) worked, in a separate conda env, running fp16 vision models without issue. Keep that env separate from your main Python, and call its python.exe directly for GPU jobs.