I’ve seen way too many people see this error and immediately go reinstall their GPU drivers or CUDA Toolkit—totally the wrong direction. The error message is pretty straightforward: the PyTorch you installed is the pure CPU version, it wasn’t compiled with CUDA code at all, so it can’t see your GPU. Has nothing to do with drivers. Nine times out of ten, some custom node’s dependency installation pulled a CPU-only torch from pip’s default source and quietly overwrote your GPU version.
Debugging is easy—just run print(torch.cuda.is_available()) in your environment. If it returns False, you got hit. Fix is simple: uninstall the current torch, then install the wheel for your CUDA version from the official PyTorch index. Works the same for portable packages and manual venvs. When picking the wheel, match it to your GPU—the new 50-series cards have extra version requirements, so don’t blindly follow old tutorials. Whole process takes three minutes, stop messing with your drivers.
